I’m making progress on my agents manual! I think I finally figured out how I want the thing to look and feel.

I’ve built and rebuilt this damn thing so many times in this process, which is actually part of the process. I am a lazy workaholic (h/t Rick Rubin) - I have to spend time in the work, even if it feels like it’s not going anywhere, until ‘suddenly’ things click.

Whilst in the process, you find yourself wanting tools to exist to make things easier for yourself—that’s a huge part of why learning agents and how to steer them is so good.

You can build tools to enable you to build things.

I spun up this tool before bed last night where I can comment/delete on copy whilst I’m building, which I copy as one big block as agent feedback.

Ben’s Bites is brought to you by Attio

Attio is the CRM for the new way of GTM. Get agents working on every account, surfacing opportunities, and handle the work that used to take your team days. Open your inbox, the follow-ups are drafted. Walk into a meeting, you're already briefed. Got a question, just Ask Attio. Start for free today.

Headlines

• Codex has two new additions: Plugins and Sites. Plugins are pre-built collections of skills, connectors to relevant apps (like Figma for designers) and instructions tuned for specific roles like data analysis and product design. Sites lets users create a shareable website/app with a database, file storage, env vars, access controls, and more. Initially only available to business and enterprise users.

• A bunch of new open models released recently -

  • Gemma 4 12B - Multimodal (i.e. accepts images and audio as input) and performs nearly as well as the two-month-old 26B variant.

  • Ideogram 4.0 - 9.3B model for image generation. Trained on JSON prompts for control over the layout, colours and text for each element on the image. Also check Reve 2.0 for the focus on layout of elements in an image (but it’s closed-source).

  • Miso One - 8B text-to-speech model claiming expressive speech with 110ms latency.

  • Also, just like Cursor’s Composer, more companies are trying out fine-tuning big open-weights models for their domain-specific work. Latest entry → Harvey got a Kimi 2.6 agent to beat Opus 4.7 on its legal benchmark at ~11x lower cost.

• Microsoft Scout is an always-on Microsoft 365 agent built on OpenClaw (reminder: openclaw is open-source). Different approach from what Google is doing with Gemini Spark.

• Ramp Stack - An accounting assistant that helps with month-end close work: reconciling accounts, preparing schedules/accruals and more with reviewable sources. They also published a nice blog post explaining their efforts to benchmark Stack against other frontier models.

• Financial fraud is evolving fast. It’s time to fight back—with AI. Read MIT Technology Review and Plaid’s report to see how technology is reshaping financial defenses. Learn more and see how smarter tools and industry collaboration can help fight against the rise of fraud. Read the report.*

Subscribe now

My feed

• Smallest AI lets you deploy voice agents at scale, powered by realtime STT & TTS and production-ready telephony infrastructure.*

• Bloom turns your brand assets, site, decks, Figma and socials into a callable system that agents can use via API/MCP to generate on-brand assets.

• Windsurf is now Devin Desktop. It manages fleets of local and cloud agents from the editor. Nous also released a desktop app for its CLI agent Hermes.

• Hallmark v1.1 - open-source design skill for coding agents.

• 40% of Cursor's internal PRs now coming from cloud agents. More deets in their post about lessons from building cloud agents.

• Factory Router routes each agent session to the right model and keeps near-frontier performance at 20-25% lower cost in its benchmarks.

• The next frontier of visual AI is code.

• ViBench - benchmark from Replit with tasks focused on end-to-end app creation; Opus 4.8 beats GPT-5.5 on price/performance for vibe coding.

• Skills for macOS - app for browsing and editing local skills, MCP configs and plugins.

• Ollie - AI assistant for parents to manage the chores to free up time for family.

• Television - visual workspace for personal agents. Notion-like kanban board vibes but with each tile attached to an agent.

• Building software is learning - it’s an iterative process that will run into questions and obstacles. You should want that to happen as fast as possible.

• A functional taxonomy of world models.

• Collection of agentic engineering hacks for June 2026.

• Modern Engineering Values - a workflow and engineering values built after shipping several mostly or fully AI-written projects.

• SDKs I’ve come across:

  • Email SDK - unified API for sending emails. works across multiple providers.

  • storagesdk - object storage with snapshots and forks.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’m spending as much time as possible off Twitter, and any other general doomscroll distractions. I’ve got work to do (the course/manual/whatever it’ll be called).

Aiming to get at least the preview ‘lessons’ out this month.

I’m not going to force any wisdom out in today’s intro because it’s not my style.

On to the condensed version of the AI madness…

Ben’s Bites is brought to you by Smallest AI

Pulse is world’s fastest speech-to-text model (#1 on Sierra’s μ-Bench for P95 latency) with top performing accuracy (under 5% WER on Artificial Analysis leaderboard), works across 39 languages and 100+ accents.
Get $25 free credits, use code BYTE25-N3CX3UKV (valid till 6/6)

Headlines

• Claude Opus 4.8 is out, with dynamic workflows in Claude Code. Claude now writes an orchestration script, then spins up subagents in parallel to work through complex tasks.

  • Dex’s take: this doesn’t prove loose multi-agent systems work. Deterministic workflows around small agent loops are more reliable.

  • Claude Opus 4.8 - Simon Willison calls it a modest but useful upgrade, mostly because it’s more honest about uncertainty and less likely to miss flaws in its own code. Every’s vibe check is more bullish: they found it a big jump from 4.7, strong at coding/writing/knowledge work, and competitive with GPT-5.5 on their internal senior-engineer benchmark. The catch is the harness: the model is back, but Claude’s app still feels messier than Codex.

    • Scores top on ARC-AGI-3, tripling 5.5’s score

    • Datacurve’s new benchmark places it below gpt 5.5 and only marginally better than 5.4. Using a lot more tokens…$$$.

      

• Anthropic filed a confidential S-1 and raised a round $65B Series H at $965B post-money. IPO this year?

• NVIDIA and Microsoft are reinventing Windows PCs for personal AI agents — RTX Spark is a 1-petaflop Windows superchip with up to 128GB unified memory, full CUDA/RTX support, local 120B model support, and new Windows agent security primitives + NVIDIA OpenShell. Microsoft’s Surface Laptop Ultra is the flagship device.

• Codex is becoming a productivity tool for everyone

• Stacker's AI Accelerator is offering $500k in inference credits to businesses ready to go AI-first. Selected companies get credits and hands-on mentoring to deploy AI agents across their operations. Applications close June 9th. Apply now.*

Subscribe now

My feed

• Stacker is an AI coworker that joins the dots in your business. Stacker lives in Slack, connects your tools, and does the work.*

• ChatGPT has a new table of contents UI for long chats, and a full-screen long-form writing mode that can save drafts to your Library.

• Codex got computer use + mobile remote control on Windows and a Python SDK.

• Replit added a Canvas to create variants, annotate, compare design directions and apply changes back into your app.

• Figma Make now works on your local code - visually edit the app, annotate, chat and open PRs.

• Linear’s new product Diffs adds PR review into Linear, with guided reviews and agent iteration.

• SkillSpector - a new security scanner for skills by NVIDIA

• Google AI Studio now lets you build apps that connect to Gmail, Drive, Sheets and more without jumping through other Google Cloud screens.

• Agent Cookie - syncs cookies, CLI tokens and API keys from your laptop to a Mac mini running OpenClaw/Hermes.

• Agent Handler for Employees: secure AI access for every employee.

• Impeccable 3.5 - design skill for coding agents with model-specific anti-pattern rules.

• Four tips to help agents understand your codebase.

• Sandboxes are becoming the OS for agents.

• 30 mins epsiode on /goal and how to use it in Codex.

• Building Grok Imagine in 3 months and Video Agents.

• The most rational take on AI you’ll hear this year

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

“SaaS may be dead” - me, on tuesday

Just signed up for another SaaS tool - me, yesterday

I’m trying really hard to make interactive components for this course/reference manual I’m making. So you as a user can feel the concepts, to help understand them.

I’ve tried so many models, tools and ways to try and develop my own component styles that look good and feel right. And I think I finally found it…

I tweeted my frustrations and Pietro, who I met at OpenAI’s Dev Day last year reminded me to try Magic Path. You can have multiple agents generating design assets, components, animations, whatever on a big shared canvas.

I gave it a go on a fun experiment first and it generated some pretty awesome mechanical-style components.

So now I have an actual workflow and tools to generate all the components I’m after. I can play with different styles and tweak the smaller parts of the components - the buttons, prompt input box, etc.

So I blew through the Magic Path free plan pretty quickly and then promptly signed up for a pro plan 😬.

Ben’s Bites is brought to you by Palabra.ai — Real-Time Voice AI Translator

9.3× cheaper than a human interpreter. Palabra.ai delivers real-time voice translation in 60+ languages for calls, events and streams – or embed into any app via API. Trusted by DHL, UNICEF, Paramount, BCG and Deloitte. Try it free.

Headlines

• Claude Code now has a security plugin that checks code as Claude writes it and warns when it spots common risky patterns, like unsafe command execution, insecure HTML handling, or dangerous Python code.

• DeepSWE tests agents on 113 original long-horizon tasks across 91 active repos and five languages. Prompts are shorter than SWE-bench Pro, but the fixes are much bigger: 668 lines and seven files on average. Current leaderboard: GPT-5.5 70%, GPT-5.4 56%, Claude Opus 4.7 54%, Claude Sonnet 4.6 32%.

• From the board to building the Software Factory. It doesn’t happen that often, but I’ve seen it a few times recently - investors in a company leave to join the company that they backed. Madison joining Factory is a big signal and a great addition to the team. If you remember, I am an investor in Factory who joined last year but I left earlier this year due to our army of young children running my life, leaving less and less time for work work. 3 under 3 is still A LOT of work 😅.

Subscribe now

My feed

• Software after software - this was a great read, highly recommended. Thorsten runs the coding harness, Amp and always has great takes on the space.

• Clanker - A word for the machine.

• Mainframe - turn the work done by your agents into short recap videos for your team.

• Granite - long-term document for all your files. Drop them in without any tagging/folders, and later search for them in plain English.

• Surya OCR 2 - 650M parameter OCR/document model.

• Claude Code trick for non-technical tasks - put a bunch of files in a folder, then tell Claude Code it can write scripts and make HTML.

• Ramp used 10,000 home-grown security agents to find, validate and patch nearly 100 security issues in six days, with humans reviewing PRs before merge.

• Slippery Slope - It’s easy to let agents get in between a person and their craft.

• Supermemory - Building blocks for adding context to your agents.

• Cursor trained Composer 2.5 by doing RL inside the actual Cursor harness.

• Polar - fine-tune a model with your agent harness as the training environment with no code changes.

• Parse 2.0 - the most accurate document parsing API in the world.

• OpenAI are slurping up a ton of talented builders, most recently, Eric who built RepoPrompt. Great get for OAI, and congrats to Eric 😊

• Auto-review skill for your agents, from the power-house shipper Peter Steinberger.

• howtoeval - the no-bullshit guide to eval’ing AI agents.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

It’s a heatwave in the UK, and I can’t think straight. Dan went on Lenny’s podcast to talk about agents, automation and SaaSpocalypse, which is worth checking out.

While I generally agree with it all, the one thing I can’t shake is the bullish case for SaaS. I think SaaS is in trouble…

I don’t think it’s in trouble because you can build your own versions with a plan, patience and prompts.

With SaaS tools, you pay monthly for a set of features and stability. The problem is that the tool needs to keep growing. They add new features, change the interface, do all sorts of things you may not need (and often hate).

The tool is for the masses but you may have only wanted a sliver. The tools can outgrow you and why you bought it in the first place.

So if I only need a sliver, I’ll go looking for pieces I can pull together myself. A document editor over here, an agent there, etc.

I think SaaS companies that can unbundle their building blocks and sell them as composable pieces for users (as well as the all-in bundle) could be a very interesting bet.

WorkOS do this well. Their tagline on Google results is literally “WorkOS is a set of building blocks for quickly adding enterprise features to your app”.

Stripe does this too.

In the age of customisable software and using tools with agents, I can’t imagine paying for a tool that I can’t change its feel or features.

It’s why I think API/CLI/SDK-first companies are in a really interesting place.

Ben’s Bites is brought to you by Sherlocq

Sherlocq is an AI-powered regulatory intelligence platform built for compliance, legal, and regulatory professionals. Research regulations, analyse documents, and run sanctions checks across 30+ jurisdictions and 320+ sanctions sources, cutting research time by up to 70%.

Headlines

• Appshots in Codex attach the active Mac window to a Codex thread with a screenshot + available text, including context beyond what is visible. Goal mode in Codex is now out of the experiment stage - it helps Codex execute multiple-step workflows with a single outcome in mind, with the ability to check back and compare its work with that goal. Plus, there are other updates like sharing plugins, using a Mac while it’s locked and more.

• MCP is getting a major upgrade (finalizes on July 28). It should be much easier to run MCP servers reliably at scale. The update also adds built-in support for app-like interfaces, long-running tasks, stronger login/security rules, better tool definitions, and a clearer process for changing the protocol over time.

• Perplexity open-sourced Bumblebee - A safety scanner for developer machines. It checks for risky packages, browser/editor extensions, and AI agent configs, without running any of the tools it is inspecting. So it can flag supply-chain risks without accidentally triggering the same risky code.

Subscribe now

My feed

• Cloudsail - fresh Cloudflare Sandboxes for coding agents, with shell, Codex and GitHub access.

• WorkOS released auth.md, an open protocol for agents to register for web services on behalf of users

• Building Pi With Pi - using agents for issue research, PR wrap-ups, context tracking and parallel work.

• An AGENTS.md file with a year of small agent-behaviour improvements.

• release.bar - trending active GitHub repos, with a dashboard for each repo/builder. (see example). GitHub should probably have shipped this already.

• Hyper - Bun API framework you copy into your codebase instead of installing as a dependency.

• Kakuna - Checklist-driven codebase hardening, and a production-ready repo skill that turns a fragile vibe-coded app into something maintainable and agent-friendly.

• Cursor’s internal /thermo-nuclear-code-quality-review skill deletes complexity, blocks huge files and rejects PRs that work but make the codebase messier.

• Simpleplayer - clean default video player with custom overlay controls.

• Running a startup solo with AI agents - OpenClaw as Chief of Staff, Codex/Devin as engineering and giving agents real email, calendar and GitHub access.

• zero2claude - free course from zero terminal experience to shipping with Claude Code.

• Find repeated workflows across recent sessions/memories, then create only the smallest useful skill, subagent or automation.

• Shades - Chrome extension that hides sensitive inputs while screen recording or sharing.

• Plannotator - annotate agent plans before they run and review the output in full diffs.

• ChatGPT for PowerPoint - create slides, ask questions across a deck, and make updates inside PowerPoint.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

My first fund is now at 5x! I had a very notable firm not invest in my latest fund as they were assessing if I was a 5x fund returner… seems like I am. Fund 2 is also at ~3x with 55% IRR. Fund 3 I’m starting to fundraise for again, looking for operators, $100k minimum. Deployed 4 cheques so far. I invest in developer tools and infrastructure - essentially anything for an agent-first world. If you or anyone you know would be interested, please let me know.

Google’s I/O event on Tuesday was overshadowed by Andrej Karpathy joining Antrhopic’s pre-training team under Nick Joseph to build and lead a new group focused on using Claude to accelerate pre-training research. Using Claude to help pre-train Claude models.

And now they have the compute… SpaceX’s IPO filing discloses Anthropic will pay $1.25 billion monthly for compute.

Just as Anthropic project $10.9 billion June quarter revenue and its first operating profit. Which could well lead them to a valuation higher than OpenAI. Who have been reported are potentially filing for an IPO imminently (some sources say as early as tomorrow) - but nothing official or confirmed.

Ben’s Bites is brought to you by Attio, the AI CRM

GTM Atlas is the map for modern go-to-market. Written by top operators, Atlas is a free resource covering the full customer journey, with systems thinking that scales with you. Curated by Attio. Mapped by operators. Read now

Headlines

• Google I/O wasn’t all that interesting.

  • They have a new model family aiming for any input/any output. Gemini Omni Flash - the first model in that family generates and edits videos. The technical difference between Omni and Veo is similar to that between Imagen and Nano Banana.

  • For general model upgrades, they only released Gemini 3.5 Flash, which, on paper, is better than 3.1 Pro. This model is fast and intelligent, but not cheap anymore. It also has the knowledge cutoff of Jan 2025, so it does not know about vibe coding and beyond. 3.5 Pro is coming next month, but I don’t have high hopes.

  • Antigravity is now a clone of Codex, Conductor and every other AI coding agent app. It’s usable — but I couldn’t find a way to bypass permission quickly. They have also made installing the IDE optional.

  • Their answer to OpenClaw is called Gemini Spark, a 24/7 personal agent that works across Workspace, but it’s “coming soon”, so no idea how well it works.

• OpenAI says one of its models solved a famous math problem, and external mathematicians checked the proof. They also added a public image verifier for images made with ChatGPT, the API and Codex. It checks C2PA metadata and also SynthID (from Google). Try it here.

• Figma has a design agent inside the canvas. It can start from a design layer, generate multiple directions in parallel, make bulk edits, use your design system, and work in the same file as your team.

• Now any product developer can ship globally: Neimo MCP turns Claude, OpenAI's Codex, and Manus into regulatory experts across 200+ jurisdictions. Built by the team behind compliance for some of the world's largest games and platforms. Try Neimo free 👉*

Subscribe now

My feed

• Handinger: Build AI agents in plain English, connect them to your tools, and kill the boring admin work nobody wants.*

• Factory’s Deferred Context Engine - Droid now loads tools more selectively to cut context size by 40%.

• Getting the most out of Codex.

• Roughdraft - local open-source interface for commenting and suggested changes on markdown docs/plans.

• DiffsHub - replace github with diffshub in a public GitHub diff URL to virtualise and inspect huge diffs quickly.

• Lapdog from Datadog - local tracing for reasoning and tool calls in Codex, Claude Code and Pi.

• Granola Briefs - searches your email, web and previous meeting notes before a meeting, then gives you three bullets.

• Taste MCP - what if your design preferences could follow you into Codex, Cursor, Claude Code?

• OpenAI Guaranteed Capacity - companies can pre-book OpenAI compute for 1-3 years, so important products and agents don't get throttled when demand spikes.

• Parallel Web Systems Index - lets publishers see which agents are reading their content and get paid. First partners include The Atlantic, Fortune, Every and Packy.

• Active Graph - open-source framework for long-running agents to remember what happened, react to new events, and compare different agent runs.

• Making computer use reliable in production.

• A poisoned third-party VS Code extension compromised a GitHub employee's device, and ~3,800 internal GitHub repos were exfiltrated.

• Cohere released Command A+ - an open-source enterprise model. supports text/image/tool use and runs on two H100s/one B200 with quantisation.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Google I/O starts today, and Logan tweeted: “The model is the product”. There have been some rumours that the latest Gemini model scores similar on benchmarks to GPT 5.5 - but we’ll see how it feels when actually using it - previous models also scored well but didn’t feel great to work with.

When models are so good, harnesses will be much less important. I just don’t think today is the day that happens. And on that point, the role of a harness will probably just shift - instead of managing how/which tools to use, the system prompt, context management etc it could be managed agents, sandboxing, cloud/local management.

I started using Codex on my phone…but not all that much to be honest. A lot of the agent harnesses these days have ways to control your sessions from your phone - Claude Code has /remote-control, Pi can build one for itself (i use a telegram one) and Droid has mobile web + Droid computers.

Most of my mobile first work at the moment is more brainstorming than building and I find myself flitting between all these options all the time.

I used to use my OpenClaw bot like an addict, but haven’t spoken to the poor bastard for weeks now.

It may help that I’m currently focused on just one (ish) main thing - this ‘course’. Which is really more of a library or reference manual on how I think about agents, how I steer them and build with them.

Ben’s Bites is brought to you by Hyperagent from Airtable

Hyperagent, the cloud agent system with full computing environments, is giving $10M in inference credits to help founders build and run agent-first companies. The first 500 qualifying applicants gain access to this limited founder offer. Applications close May 31st.

Headlines

• Codex now connects your Mac to your phone. You can start tasks in Codex from your phone, but the actual work still runs on your Mac, devbox or remote machine, i.e. files, setup and credentials stay where they are, while you can approve commands, answer questions, and review diffs from your phone. This update also brings Hooks to Codex.

• Anthropic is acquiring Stainless, a platform to build SDKs (also used by OpenAI), and they are shutting the service down. Also, at their London conference, they added self-hosted sandboxes and MCP tunnels to Claude Managed Agents - their “running agents made easy” product for companies.

• Cloudflare tested Anthropic’s Mythos against 50 of its repos. Quick takeaways:

  • Mythos is great at spotting real attacks, which are often many small vulnerabilities connected in a chain.

  • A single model, however smart, without a good harness leaves a lot to be found.

  • “Find bugs fast and patch them faster” is not a good idea. Teams need to focus on making bugs harder to chain (even if they exist) and to exploit.

• Cursor’s Composer 2.5 (partly trained on SpaceX’s GPUs) is out. The selective benchmarks that Cursor reports put the model roughly at the same place as Opus 4.7-xhigh and GPT-5.5-high, while being much cheaper than them.

Subscribe now

My feed

• Two AI startups worth watching: Magicpath (design canvas) and Raindrop AI (monitoring agents in production), both of which are making their products usable by external coding agents like Claude Code or Codex.

• Even Grok/xAI has a coding CLI now. Let’s see what Google does with Gemini CLI at I/O today.

• Linear Agent can now read the codebase directly to build a hypothesis, investigate support questions, find people who worked on a feature, and more.

• Best practices for running Claude Code at scale.

• Citadel’s founder, Ken Griffin, one of the anti-AI hype people, is now saying that they are seeing high-skilled jobs being “automated” by AI.

• Browse.sh from Browserbase - open-source catalogue of skills/playbooks for agents to perform tasks on the internet.

• Watchmen - skill files your coding agents should already have from your past sessions. Local and open-source.

• Devin Auto-Triage monitors bugs, alerts and incidents, investigates them and comes back with context, next steps or a PR.

• Motus Tracing - open-source observability for AI agents.

• designmd.sh - a public registry for DESIGN.md files, so agents can understand design systems from repos.

• Jason Liu on Codex maxxing - daily primitives for durable threads, shared memory, and keeping Codex useful across a real workflow.

• Taste MCP beta - portable design preferences for Codex, Cursor, Claude Code, etc.

• Claire Vo and Thariq on “HTML is the new markdown” - using HTML artifacts as specs, micro-UIs, and human-readable agent context.

• Brian Lovin’s Notion Worker - syncs the people you follow on X into a Notion DB with optional AI enrichment.

• Benedict Evans’ new “AI Is Eating The World” deck.

• Coatue says its AI framework moved from “follow the GPU” to “follow the gigawatt”.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

When an agent asks for feedback it feels like the levels are

1. type your response

2. voice-to-text your response

3. + images to your feedback

4. get the agent to use the browser

But I just started screen-recording and talking then giving that file to my agent

This is me, in droid, like 30 mins ago. It pulls together a pretty great visual report you can easily review. I can navigate to other websites or apps and show what good looks like from other people, I can highlight specific points and it’ll recreate those points with GIFs.

It gives itself an ‘actions’ checklist underneath. And just feels great to have screenshot → my feedback → action for the agent.

It’s pretty great so far, and then I’ve got these html files saved in my projects to always refer back to - will be good for a build log too.

Probably not great for the token conscious out there - and thinking about it, I could probably use ffmpeg to create actual clips of the video if I wanted. Agents read frames well though so it’d be more for me if I did.

I turned it into a simple skill:

Let me know any cool use-cases or remixes of this 😊

Ben’s Bites is brought to you by Hyperagent from Airtable

Hyperagent, the cloud agent system with full computing environments, is giving $10M in inference credits to help founders build and run agent-first companies. The first 500 qualifying applicants gain access to this limited founder offer. Applications close May 31st.

Headlines

• Your Claude plan is changing if you use third-party tools (like Conductor, Zed, Openclaw, T3 Code, etc.) with it.

  • Separate limit for all such usage. Provided as extra monthly credits equal to the value of your plan.

  • No subsidised tokens, credits won’t roll over and usage after you burn through these credits is billed at API rates.

  • Using Claude in Claude Code, Claude app, etc., stays the same and is separate from this.

  • Starts from June 15th, but they are increasing your weekly rate limits by 50% for the next two months.

• Google announced some Gemini on Android updates before I/O - add features like auto-completing forms, rambling voice notes to clean text, and some app automations under the name “Gemini Intelligence”. They also announced a new class of laptops called Googlebooks, not to be confused with Google Books.

• Notion has a developer platform now. The biggest addition is a markdown API. Also, devs can sync outside data into Notion, build tools for Notion Agents, run code on Notion’s infra, and eventually bring agents like Claude/Codex into Notion as teammates. But I think people who don’t call them developers will use this.

  • They also launched a CLI called ntn.

• Vercel published an AI Gateway production index based on real usage across apps and agents. Anthropic leads spend (61% — due to opus), Google leads token volume (38% — due to flash), and agentic workloads are 59% of token usage. Most large teams route across many models instead of betting on one lab.

Subscribe now

My feed

• Cursor now lets you run cloud agents inside a fully configured development environment.

• Orca - Claude Code’s agent view but for Codex, OpenCode, Droid and Pi.

• Oboe - LLMs wrapped in a way that helps you learn.

• Interfaces.dev - A monthly design engineering magazine about building great interfaces.

• Anthropic CFO Krishna Rao on compute allocation, pricing dynamics and model company economics:

• AI IQ - frontier AI models, scored on the human IQ scale.

• Intercom is rebranding the entire company to Fin, their popular AI agent.

• Executor - Convert MCPs/OpenAPIs servers into code mode under the hood, 100% local on your device.

• How OpenAI built a safe sandbox for Windows.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I watched this this morning; ‘Agentic coding is a trap — and we all fell for it’ - it’s surprising relevant past just developers.

enjoyed this because what i took away;
learn the system

to me, this is the difference of vibe coding and agentic engineering. i'm actively trying to learn the system, not the syntax.

syntax is what i couldn't grapple when attempting to learn to code. but the system is clicking more for me the more i build

im miles away from a 'competent software engineer' but im only building things for myself, so i dont *need* to be - but the more i build, the more that clicks into place.

i didnt realise it when i was slinging no-code in 2018 but it was a version of learning parts of a system to get software to work (webflow - frontend, airtable - database, zapier - api/backend). it had limitations but now i replaced all of that with code.

having actual competent engineers create skills and systems to help a sloppy codebase or process is helping (as well as better models). but i do rely on that for years i didnt spend learning to code.

stay curious folks

tweeted it earlier

Link to the original blog post.

And I’ve been tweaking my AGENTS.md + setup files - as I’m getting closer with the ‘course’ I’m reminding myself to not let your docs slip.

Oh and…

Ben’s Bites is brought to you by Lightfield

Lightfield is an AI-native CRM that just shipped Skills. Define any workflow in plain English and trigger it with a sentence. The AI agent executes against your full CRM data graph with code execution, web search, and file I/O. 3,000+ startups on the platform. Try free

Headlines

• You can now work with all your Claude Code agents in a single window inside the terminal. You can see their status and reply inline to unblock when they need your input. Any running session can be moved to the agent view with /bg.

• Codex now works directly in Chrome on macOS and Windows. It can use sites and apps across tabs in the background without taking over your browser.

• OpenAI also released three new Realtime models in the API: Realtime 2 for voice-to-voice use cases with best intelligence, Realtime Translate for audio translation across 70 input and 13 output languages, and Realtime-Whisper for live speech to text.

• OpenAI released a cyber defence product, Daybreak. Their OpenAI’s answer to Anthropic’s Mythos?

• Thinking Machines finally have a model to show us (not letting us try though). They are calling them interaction models. Basically models where you can chat with audio and video input with audio outputs. It seems really impressive for the capabilites they are claiming, for example, time awareness, simultaneous speech and visual cues, but all similar products (ChatGPT’s Advanced Voice Mode, Gemini Live) fail when put in users hands.

• OpenAI is starting a deployment company in partnership with major consulting firms. It acquired a 150-person AI consulting company, “Tomoro”, to set this up and is putting in $4B of initial investment. The goal is to work with other companies and build AI systems for them.

  • I think this means they’re going to effectively transform a ton of knowledge workers and upskill them to knowing how to work with agents. ie able to be a builder. And if you’re a builder → you can use Codex. You can see how it all links 😊

Subscribe now

My feed

• Artificial Analysis is testing rank models + harness combinations together in their Coding Agent Index. Among the combinations they have tested, Opus 4.7 with Cursor CLI is on top with GPT-5.5 in Codex and Opus 4.7 in Claude Code at a close second.

• Ramp trained a small RL model with Fast Ask with Prime Intellect for spreadsheet Q&A. They say it beats Opus by 4% on exact match accuracy at Haiku latency.

• Replit Parallel Agents lets Replit Agent break work into tasks, run them in isolated copies of your app, and merge them back after review.

• Notion Skills - Brian Lovin is using a Notion database like an app store for agent skills, with two-way sync to Claude, Codex and other local agents.

• React Doctor v2 catches bad React code from agents.

• Printing Press - generate agent-native CLIs for apps like Linear, ESPN, Kayak, etc.

• The Claude Platform on AWS is now generally available. AWS customers get Claude API features with AWS auth, billing and commitment retirement.

• OpenAI’s API has a new Files SDK for object and blob storage and an OpenAI Developers plugin for Codex to build faster with OpenAI APIs.

• Parallel AI’s Monitor API is now GA. It sends web push updates to background agents instead of agents constantly polling for changes.

• zero-native - build native desktop and mobile apps with web UI.

• A spec for how interfaces should present Markdown.

• 7 Powers in the age of AI for building a company.

• a framework for hackable software i.e. apps that ship with their raw source code, where users can modify them using coding agents.

• New research from Anthropic translates the inner workings of Claude into text and teaches it good behaviour using fictional stories.

• Peekaboo 3.0 - Peter’s macOS computer-use tool got action-first automation, unified screenshot + UI detection, cleaner JSON and better snapshots.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

And at the bottom I’ll touch on the tools I’m using day to day.

Subscribe now

What did I build this week?

An email app…

I use Gmail. I’ve used Superhuman for years. I like it a lot. It is fast, keyboard-first, clean, and is good software. But like many saas products, it keeps adding features that I don’t need and more importantly, I don’t need to be paying for email.

I wanted a split inbox and rules to organize my emails.

Kicking off with Codex:

The end product has split inboxes, shortcuts, command palette, reply/compose, 20 second undo send, one-click unsubscribe, search, rules, filter sync, email rendering (mostly 😬) and built in a way that an agent can use it natively.

The code probably has slop in it, I wasn’t planning on sharing it but others have asked. So I will. But I’m only building it for myself.

Version 1

I already have labels I use all the time:

• investing - for porfolio companies and LPs

• fyi - receipts, updates, info I never need to read

• cal - calendar invites

• news - newsletters

• pitch - PR emails that I don’t need to read

I don’t need AI to read every email and categorise it for me - filtering by domain or email address does a great job.

Codex did the first pass but struggled with the UX/UI so I moved to Factory which handled the polish (I could flit between Opus and GPT 5.5), many fixes, testing and getting the product right (enough).

First real issue

It was laggy. I’m using the Google Workspace CLI which can authenticate, fetch emails, and update labels - all the actions I need basically. The whole app is built on top of this.

Investigate why there is a delay with anything in this app. I want everything to feel instant.

It found the app was pinging Gmail too much, too often.

So it then started showing cached data immediately, refreshing in the background, prefetching thread details and updating the UI optimistically when I archive or label something.

This is one of the reasons I like building with agents. I’m learning along the way - obviously a database would help here - so we added that.

Labels and rules

At first labels were just labels. Press L, pick one, done.

If I label something investing, it should move to Investing. If a rule says a label skips the inbox, Gmail should remove it from the inbox.

It’s obvious with Gmail filters but my agent was defaulting to a local rule file and not actually syncing with Gmail. Factory sorted it out thankfully.

Then I decided I wanted to make rules richer - when I add a label rule, I should be able to define if it should happen to all domains or just that specific email address.

I originally did that with a notification to choose after labelling but then I decided to make it part of the label modal. Code is cheap - try it, then if it feels wrong, change it.

Making it agent-friendly

Adding reply functionality

Then I wanted to be able to reply, obviously.

This was a bigger jump than I thought… but after a bunch of UX iterations in Factory it feels pretty good.

Email rendering

Showing email sounds easy until you try to show email.

Plain text, Gmail HTML, newsletters, receipts, calendar invites, inline images, signatures, quoted replies, tracking pixels, weird MIME stuff. Horrible little world.

Codex improved the system but with a bunch of specific rules like if this do that.

So I stepped away and tried to think about other ways to approach it. I know Obsidian has a web clipper that displays nice previews of web pages. So I got Codex to reverse engineer the chrome extension to understand how it works.

So then I asked to apply that approach to email rendering. It’s still not prefect and I’ve given up on perfect for now so normal emails look good (enough) and html heavy emails (like this newsletter, amazon, etc) just get shown as they are.

But some emails have stupid signatures that look like designed HTML emails. So I got Factory to fix those and essentially hide that part of the email from the user.

Factory then found one of my favourite bugs. A normal email with a fancy signature rendered like a designed newsletter.

Why? Logo/table signature. The classifier thought the signature meant the whole email was a designed HTML email.

Search and All Mail

Then I added a search page for all my emails.

And then, All Mail, because sometimes I do need to see what’s been archived.

But it was a bit off…

The fix was to lazy-load All Mail and stop hammering Gmail.

My takeaways

I don’t advise building an email client 😅.

If you want a feature (and then another) just build it and see what feels right. Code is cheap.

Switch between tools (harnesses) and models.

Think about other tools or apps that may do all or parts of what you’re trying to do. Ask your agent to reverse engineer it.

Use fresh sessions outside of your project to get unbiased opinions on implementations. Your agent is influenced by everything in it’s context.

Ask the agent a lot of “I assumed it was doing this, is it?”, “Why does this happen?”, “Whats the best way to do this?”, “What are the tradeoffs of this approach?”

Get agents to use the browser to test things out, a lot.

You learn a lot from building for fun or for the sake of it. So just build stuff!

My stack

So many tools. So little time.

Everyone feels the same - so just find tools you need for the job you need to complete. Don’t worry about all the new shiny tools every day.

Day-to-day ‘work’ i.e. brainstorming, talking about ideas, essentially everything non-coding heavy (like bigger projects or apps) I’m still using Pi, but now with GPT 5.5 since Anthropic has cracked down…boo! But I’m liking 5.5 - GPT models tend to follow instructions more directly so it really matters what you have in your AGENTS.md and context.

I’ve also spent the last 2 weeks in the Codex app - it’s really good. Use a VPN if in EU/UK so you can use the chrome and computer use tools. But I still find myself preferring the terminal (with Pi) because you can see everything the agent is doing and jump in to course-correct (or tweak your instructions). But IMO this app is better than the Claude variants. Don’t bother using Cowork - it’s limited and they’ll merge it into Claude Code soon enough - don’t be put off by ‘code’ in the name. All these coding agents are just great general agents.

For ‘real’ coding i.e. when I’m serious about building something, I use Factory. It’s the best harness for code, especially if you don’t know how to actually code. It’s more complete, and seems to handle everything much more than others I’ve used - plus you can switch between Claude and Opus models very easily. They recently launched a $100 plan too with better limits too.

For full stack apps I’m sticking with Vercel, Stripe and Supabase. I just started using Stripe’s projects.dev which will set up all three of those apps (and many others) all in one. Your agent can basically just use projects.dev, set up, manage and link all the other services you need to get apps setup properly. Super helpful as you don’t have to manually add or manage each yourself.

I use here.now a lot - I’m always spinning up new sites for random ideas or tasks like visualising data and things of that nature. Plus sites are easier to read than pure documents. Just give the instructions (copy from the homepage) to your agent and you get free websites spun up on-demand. Super slick and easy.

For document editing and writing I’ve been using Clearly quite a lot recently. I like it - I don’t know what I really want out of a document editor like this, truly I just want to view and edit docs within whichever tool I’m using (pi/droid/codex) but none let you do all of that…yet.

That’s mostly it - other than a bunch of CLI tools like downloading youtube videos or podcasts, getting transcripts, and I use markdown.new for my agents to easily grab website information (It’s in my AGENTS.md so my agents use it - i.e. https://markdown.new/www.example.com - but I recently discovered defuddle.md which I may switch to).

Chat with me

If you know a builder that’d find this useful, feel free to forward to them.

Have a great weekend!

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

I’m a professional procrastinator — I need to ship this course and re-write my fundraising deck for fund II, buuuut yesterday I finally built something I’ve wanted for a while.

I loved Superhuman, I used it for many years ($40/mo!) but all I really loved was split inboxes based on labels, and how quick and nice it is to use.

I don’t need (…yet? ever?) AI to read all my emails, draft replies, chase me to do things and be a PA for me, I’ve had PA’s and I always let them go.

I just want email rules - if this address is labelled ‘pitch’ (for PR pitches), archive it. If no label - it’s ‘important’ and needs a reply for me, if ‘investing’ it’s from an LP or portfolio founder, if ‘newsletter’ archive it (don’t do this or you wont see these! 😊).

Gmail has filters and labels but it’s limited and just can’t give me the UI I want to work in. So naturally, I built my own. I’ll send a ‘Ben’s Builds’ email on Saturday with more details on what I actually did. It took me ~2 hours for the first version.

I was ‘pushed’ to this by seeing Dan Shipper’s Codex-native email workflow - but again, I don’t need most of the stuff he’s doing. Email is very different for everyone. But now I can completely customise my experience…If I want daily briefs to summarise all my newsletters - my agent can do it, if I want automated actions - my agent can, and so on.

So if you send me an email, it is me who will reply - but my agent may have you labelled and organised to make it easier for me to respond.

Personal software can be built by anyone!

Ben’s Bites is brought to you by Attio

GTM Atlas is the map for modern go-to-market.

Written by top operators, Atlas is a free resource covering the full customer journey, with systems thinking that scales with you.

Curated by Attio. Mapped by operators.

Read now

Fun fact: Attio’s founder, Nicolas was interviewed by me in Sept 2020 for my previous company’s podcast. Attio is very good software!! I should’ve invested at the time 😩. But for now… I’m going to build something with it 😈

Headlines

• Free users in ChatGPT are now on “GPT-5.5 Instant” - a new model that replaces GPT-5.3 Instant. It’s significantly better at vision, understanding PDFs, web search and using your memories and past chats smartly. Its responses are also shorter in general with less emojis. It also hallucinates 52.5% less than the previous model on high-stakes prompts.

  Though recently I’ve been recommending Codex to friends with free plans of ChatGPT a lot. Yep, Codex is available on free plans. It takes them some time to understand the concept of how reading-writing a file on computer unlocks much more capability but in each case, they have come back after a day or two saying “we’re addicted to using codex”.

  — Keshav

  • Also, ChatGPT now works inside Excel and Google Sheets to build sheets, analyse tabs, write formulas and make approved edits in place.

• You can now use twice as much of Claude on all paid plans. How? Anthropic signed a deal to use all of SpaceX’s Colossus 1 data centre. (I guess no one needs/uses Grok)

• Code with Claude was a bit meh! The only new launch they did was introducing some features in Claude Managed Agents -

  • Dreaming - Review past chats and save memories from them.

  • Outcomes - Describe what success looks like, and a grader will judge the agent’s work.

  • Multi-agent orchestration - Let a lead agent break the job into pieces and delegate to specialist subagents.

• Posthog is building a code editor. Not literally, but they are making a Codex-like app that uses the data (like product usage patterns, bugs observed, errors in logs etc.) as the primary signal to code/build stuff. Here’s how they are thinking about the self-driving product loop.

Subscribe now

My feed

• Gravitee makes APIs agent-ready, helping teams govern APIs, events, and AI Agents while reducing silos and cost.*

• Skills by Entire to teach agents to explain code, search prior session context, investigate why a change happened and hand work off between agents.

• pookie - Slack helper to search messages across your workspace. It also generates memes, and connect to tools like Linear, GitHub and Stripe.

• The lines between vibe coding and agentic engineering are starting to blur.

• Clicky can now click, save ideas/links/inspo and run Gmail, Calendar and Drive by voice.

• deepsec - security harness for finding vulnerabilities in your codebase.

• Raindrop Triage - an agent to debug your agents already in production. Also works via MCP.

• Prime Intellect Lab lets you fine-tune your own models ranging from 1B to 400B params.

• How we improved agentic search.

• Everyone should have an OPINIONS.md

• Gemini API’s File Search can now search over images & audio i.e. finding 2-3 relevant images from big folder based on what’s in the image (not its name).

• @supabase/server - public beta package for server-side auth verification, client setup and request context across Edge Functions, Cloudflare Workers, Hono and Bun.

• Anthropic released 10 finance agent templates for pitchbooks, KYC screening, valuation reviews, month-end close and more. They run as Claude Cowork/Claude Code plugins or Managed Agents cookbooks.

• The “AI Job Apocalypse” is a complete fantasy.

• The artistry of text-to-speech models.

• Dharmesh says HubSpot’s goal is full API parity with the UI: agents can run on HubSpot, and agents can run HubSpot. More headless SaaS / AUX energy.

• How to use Codex for knowledge work

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I spent most of the bank holiday weekend offline for a change whilst at a wedding in the English countryside. I would’ve been more online if Codex had a mobile app but still waiting on that… this morning I did just install this skill that lets you iMessage Codex which is pretty great - essentially keeps a thread open in the app that you can message. Just paste the link in Codex and it’ll guide you through everything.

Ben’s Bites is brought to you by Gravitee

As AI Agents connect to more APIs, security risk gets harder to manage. Gravitee helps teams govern APIs, events, and AI Agents while reducing silos and cost. See what enterprise teams are prioritizing in the State of AI Agent Security report.

Headlines

• OpenAI wants non-technical users to use Codex. They are making it easy for you to switch to Codex. You can now import settings, plugins, agents, project configuration and more into Codex (from tools like Claude Cowork). They are directly improving features related to everyday work, like creating slides/sheets, plus friendlier UI changes.

• Grok 4.3 is out in the API. 1M context, text + image input, reasoning and a December 2025 knowledge cutoff. It’s priced $1.25/$2.50 per million input/output tokens, i.e. much cheaper than Sonnet 4.6 for a relatively similar performance.

• Entire, the company by GitHub’s ex-CEO, released two new things: git-sync - a utility to mirror git repos from a source to a target without needing to clone it locally and Dispatches - a feature on their web platform to generate release notes from recent ships, commits, and agent sessions by repo/date range.

• Charity Majors and Christine Yen headline Honeycomb's Innovation Week (May 12–14), a 3-day virtual event addressing observability for the agent era. Learn how the most forward-thinking engineering teams are rising to meet this challenge. Register now.*

Subscribe now

My feed

• Lightfield - AI-native CRM that learns how you sell. Describe any workflow in English, your CRM runs it on command. 3 mo free w/ BENSBITEST13*

• Sauna - learns how you work, remembers everything that matters, and actions on it (portfolio company!)

• Shared Brain by Zapier - Collective knowledge vault for your team and a personal assistant to complete tasks. Now in early access.

• Manus Cloud Computer - always-on cloud machine for Manus so bots, scripts, databases and scheduled jobs keep running when your laptop is off. Files and installed tools persist across sessions.

• Proxyuser - test all the core flows of your app via a synthetic user with a real browser, including signups.

• Web UI Bench - Same UI components built by 20 models, shown side-by-side. GPT-5.5 uses too much bland text in the UI when an icon or control is self-explanatory (compared to Opus 4.7).

• Flue - TypeScript framework for building Claude Code-style agents.

• deepsec - open-source security harness from Vercel for finding vulnerabilities in your codebase with coding agents.

• localterm - run a terminal in your browser with npx localterm@latest start.

• open-slide - slide framework built for agents. Visual edits, comments, assets and agent-readable slide structure.

• Refero Styles - 2,000+ DESIGN.md files from real products that your agent can use for style references.

• How OpenAI delivers low-latency voice AI at scale.

• crabbox - run your dirty worktrees in a remote sandbox easily. (tweet)

• OpenAI has a new opt-in feature for Advanced Account Security in ChatGPT/Codex.

• Base44’s Frustration Meter - usage-based model benchmark. Base44 says Opus 4.7 caused 43% more frustration than Opus 4.6.

• Cofounder 2 - another “run a company with agents” product that’s a combo of vibe coding, finding leads and sending sales emails.

• How Posthog plans to change in the AI era.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I learned a lot about hard work from him. Too much to put here - he transformed his law firm several times over 4 decades and most recently forced the company to adopt AI (I may have played a part here 😉).

I’ve saved the soppy shit for a letter I wrote him today - as he wrote me one when I sold my first company.

Happy retirement day, dad. I know you’re reading this - my biggest fan (the feelings mutual)

*retirement for him = started a new company, on the board of a university, school and Cardiff business council (plus whatever work I start giving him 😈)

Ok, AI stuff…

I’m becoming Codex-pilled… I’ve been using a terminal every day since realising it’s not scary any more, as it’s just talking to AI, but I forced myself to try the Codex app for tasks and I actually am starting to really really like it.

I still prefer being able to easily have multiple tabs next to one another but I like the Chat | Files (or browser) view a lot - something I never had with a terminal.

I built a dinosaur jumping game for Max and Arabella this morning. Set up a Gmail learning system to understand how to label + archive my emails (plus an automation for it). I’ve revamped my memory system and have been pilling in tons of bookmarks/youtube transcripts etc that I can reference any time.

I have a main folder ‘bites’ which has my general instructions for day-to day stuff:

So it speaks to me in a certain way, knows my memories, how I want it to behave etc. I have a running ‘todos’ thread that I work through each day - with a corresponding ‘TODOS.md’ file.

I desperately am waiting for a mobile version…

Working with agents is just files and access. I’m excited to keep trying this out for day-to-day and coding. But for ‘proper’ coding work I still use droid in the terminal.

Ben’s Bites is brought to you by Ai4 2026

Don’t miss Ai4 2026 — America’s largest AI conference! Join 12,000+ attendees, 1,000+ speakers, & 400+ exhibits in Las Vegas, August 4-6. Get case studies from Fortune 500 execs on AI Agents, Generative AI & more. Keynotes from Geoffrey Hinton, Fei-Fei Li, and Andrew Ng. Learn more & register now!

Headlines

• Cloudflare now lets agents create accounts, buy domains and deploy. An agent can provision a Cloudflare account, start a paid subscription, register a domain, get an API token and deploy without you doing dashboard/admin gymnastics. Humans still approve terms and permissions, but the setup maze is getting agent-readable. Lots of companies are moving this way, letting less-technical people like us just get stuff built without worrying about all the services in the middle.

• Stripe announced 288 things at Sessions, relevant ones: Agentic Commerce Suite, Link's wallet for agents, streaming payments and agent-ready Treasury accounts. There is also a Link CLI that gives agents one-time-use payment credentials from your Link wallet, without exposing your real card details.

• Cursor SDK - build agents using the same runtime, harness and models that power Cursor.

• Warp is now open-source and OpenAI is the founding sponsor for the repo. Warp's bet is that contributors bring taste/direction/verification while agents do more of the implementation work.

• Lightfield is an AI-native CRM that just shipped Skills. Define any workflow in plain English and trigger it with a sentence. The AI agent executes against your full data graph with code execution, web search, and file I/O. Try Lightfield (use code BENSBITESS23 for 3 months free)*

Subscribe now

My feed

• The most fun I’ve had building apps: GPT-5.5 + GPT-Image-2

• Codex masterclass workshop from AI Engineer London, plus GPT-5.5 prompt guidance

• Building Pi, and what makes self-modifying software so fascinating

• Fred Wilson on 40 years in venture — and why USV is automating itself

• Martin Fowler’s updated AI coding notes - verification, guardrails and the difference between vibe coding and agentic engineering.

• Poolside’s first public models, Laguna M.1 and Laguna XS.2, are now on OpenRouter. Built for agentic coding and long-horizon work.

• here.now now has private cloud storage for agents as well as instant web publishing. I synced all my local memory and files (2.5k files to this with one prompt - tweet is below)

• Stripe emulator skill - offline, stateful Stripe emulation for checkout pages, webhooks and deterministic CI without secrets.

• Sandcastle - Matt Pocock open-sourced his software factory.

• Quick by Amazon connects to email, calendar, Slack and local files to flag important stuff, summarize information and automate tasks.

• Some good posts on AI coding for senior engineers, harness engineering, why people hate AI, and yet, the race to adopt it.

• OSSCAR - a quarterly index of the fastest-growing open-source orgs from Supabase and >commit.

• Before GitHub - Armin Ronacher on how open-source collaboration worked before GitHub.

• ElevenMusic - discover, remix, create and earn from music built on ElevenLabs’ music model.

• OpenAI explained the goblins. OpenAI models have been increasingly mentioning creators in chats. They found out why, but GPT-5.5 was trained before the fix.

• Andrej Karpathy at Sequoia AI Ascent on vibe coding vs agentic engineering. And also linking Demis’ talk from the event too.

• Ghostty is leaving GitHub, with Mitchell Hashimoto writing up why.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I spent most of last week in San Francisco, but mostly sat in my hotel obsessing over my Stanford talk while trying to work out where I’d find the specific colour presents my kids had requested. Arabella wanted purple. Max wanted green.

Jenni is the main reason I went to SF. She invited me to Stanford and it was such a privilege. Plus, she’s brilliant. Sharp, high-agency, but warm and very personable. She also runs an AI consultancy for CEOs and gets my highest recommendation. If you get the chance to work with her, take it.

SF has this energy - how do I go faster, grow bigger, become more successful?

I get it. I like that energy. But I also came home thinking that my own version of ambition looks a bit different.

Yesterday I finally read the profile of Lenny from Lenny’s Newsletter and a lot of it stuck with me. Because it was about building a life where you can care deeply about the work without letting the work eat everything else.

That felt so ‘me’ I had to text him and let him know!

I spent an unreasonable amount of time on my Stanford talk. Not because I didn’t know what I wanted to say, but because I cared about how it landed and was presented. And also because I was trying to make the course and the talk at the same time.

Working on it also made it clearer to me that I sit in this weird middle ground where non-technical people think I’m technical, and developers don’t. But I think a lot of you are somewhere in that messy middle too. Truth is, it’s all a bit messy - everyone is still figuring out agents and AI.

Working with agents isn’t about becoming a developer, but understanding the shape of things; files, tools, systems. With a thick dollop of taste on top.

If you can steer agents, you can get technical pretty quickly. I did. You just have to be realistic that you’ll hit tons of bumps in the road. Your job is to use your agent to figure it out - and you’ll probably learn something that’ll come up again and again.

That’s what I want Ben’s Bites to be more about. Taking you along my exploration: what I’m seeing, what I’m trying, how I’m thinking about it, what bumps I ran into, come along and try whatever sounds useful. Too much ‘education’ out there is just thirsty growth-hacks to sell you something.

I’ve sold a company. I’ve got three tiny kids at home. I want to do excellent work, make good money, back great companies, and build something useful without accidentally creating a job I don’t want.

There are big, valuable companies that want me to work with them. They’d give me clout, access, money. But I hesitate because I don’t know if I can give that kind of thing my all.

Even with my fund. I’m invested in funds who have more money, better process, bigger pipelines. And yet my funds are outperforming.

But I struggle to fundraise because my story isn’t presented cleanly and my process doesn’t look like a workflow. I tinker. I talk to developers. I try tools. I see what people click. I back founders building things I think will matter.

It’s all connected. The newsletter gives me a read on what builders care about. The fund lets me back the tools that might become important. The course/workshop stuff is me trying to teach the shift I’m living through myself.

Devtools built for developers today become the tools agents use. Humans steering, agents operating. They’ll pick up the tools, compose them, run the commands, change the files, connect the systems.

We just gotta understand enough of the shape of the work to direct it well.

A lot of what I do comes down to feel. Caring about the thing enough to make it good, but not needing to turn every good thing into a machine. And squeezing every last drop for growth’s sake.

So I came back from SF without a plan to scale. Mostly I came back thinking, I want to keep building for this new class of builders. People who are curious, increasingly technical, and trying to use AI to become more capable.

That feels like a good place to spend my time. Raising the floor, not the ceiling (h/t Jenni/Jen’s Bites).

Exploring, tinkering and teaching.

p.s. the kids got their toys, and I’ll continue working on the ‘course’ with care 😊

p.p.s. speaking of care, my brother Adam just launched Hono UI - a UI kit (like Shadcn) but for projects that use Hono. Proud of him! Do me a favour and blow up his launch post 🙏

Ben’s Bites is brought to you by Attio, the AI CRM

Honestly, no one gets excited about a CRM. But then they try Attio. It connects to Claude Code and n8n through its MCP server, completely bridging the gap between my customer data and apps. Wait, there's more, like flagging churn risk and turning customer feedback into Linear projects. Try it now.

Headlines

• OpenAI released GPT-5.5 - On paper, it’s a much better model than Opus 4.7 (minus frontend design - but using their new image gen model and asking 5.5 to turn that into code is a great workaround). This comes with a price increase making it 2x more expensive than GPT-5.4 and even slightly more than Opus 4.7 on a per-token basis. But claims it’s 40% more token efficient, so per-task cost doesn't change much (Ramp reported similar results).

  • It’s become my default model since Anthropic cracked down on usage outside of their own harness. I’m in ‘thinking:low’ camp, super smart and fast.

  • I’m also trying to use the codex desktop app over a terminal and there’s lots to like - but I’m just desperate for cloud/mobile use too.

  • One more thing… it’s clear OpenAI is focusing on builders, they’re shipping like crazy, are very vocal on X, responding to users and resetting rate limits. We love to see it.

• Claude got worse. It’s official. Anthropic posted a breakdown of what and why it happened. Basically, changing the default thinking mode and system prompt changes led to a combined fall in the quality of Claude's responses, especially in Claude Code. They still clarify that they don’t switch to a quantised or “worse” model.

  • Memory on Claude Managed Agents is now in public beta. Context: Managed Agents is Claude’s API product for companies to easily plug in claude code-like agents into their products without managing the hassle of infrastructure.

• Cursor needs compute and a model. XAI needs users. And competition is higher than ever for coding agents. Cursor was raising $2Bn at $50Bn but has been paused because they struck a deal with SpaceX/XAI that gives SpaceX the option to buy Cursor for $60B later in 2026, or instead pay $10B for the collaboration. It’s a weird structure I’ve not seen before - so we will see what happens. Never a dull day.

• Your CRM knows nothing because you never update it. Lightfield captures every email, call, and meeting automatically — then lets you ask anything. “Why do we keep losing to [competitor]?” Real answers from real conversations. Get 3 months off with code BENSBITESS13.*

Subscribe now

My feed

• QuickCompare helps you choose the right LLM for your use case. Compare 50+ models on your own data in minutes with Trismik. Try free now!*

• Flipbook - every pixel on your screen, streamed live directly from a model. Another similar demo.

• Clicky lets you talk to AI and spawn an agent on your Mac. Farza built and open-sourced a version of this a few days ago, but the new clicky is closed source. OpenAI already had a version of this which you can clone.

• Things I learned working at OpenAI.

• AI isn’t replacing anyone.

• Exa for Claude - plugin to give your Claude access to websites, people, companies, and more.

• Ora.run - scan and rank how well agents can find and use your business.

• Tolaria is another app to read/write markdown files on your desktop with second-brain-esque features.

• Reelful - from a camera roll to a finished reel in just 10 minutes.

• trunks by layerbrain - turn your own storage into git remote with a minimal CLI.

• webpull - instantly pull any website into a directory of clean, searchable markdown files.

• slacrawl - cli terminal app for slack with sqlite backend.

• When and how to use MCP for apps built for production.

• create-agent-tui - skill for building your own agent harness + terminal UI.

• Aman (ex-Harvey) built an offline AI app to track his skin’s response to different diets and products with local models. (testflight) (how he built it)

• OpenAI is updating its partnership with Microsoft. Separately, Sam and Greg did an interview recently (their first one together?)

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

For a few months, it felt like Google had won the image generation space. But OpenAI is back in the game. ChatGPT Images 2.0 is miles ahead of anything. It’s beyond impressive at text, I haven’t seen any generation with typos, even with hundreds of words per image. See this example I created:

It’s also really good at creating realistic pictures, like this one of Professor Ben.

Oh, sorry, that one’s real. Ben was at Stanford this Wednesday, teaching how to build with AI agents.

Image generation is also available in the Codex app as a skill. Use it with thinking models to get the best results—that lets it think and use code/tool calls (like creating a QR from a link, searching logos from the web) and then use them as reference images. It can also create images, reflect on them and improve the generation.

People are creating realistic UI screenshots, multi-page illustrated magazines, personal style recommendations and creative QR codes using the new model.

The “generate UI as image” bit is interesting. Maybe there’s finally a solution to GPT-5.4’s lack of design taste. The latest coding models are fairly good at turning screenshots into code, but there are still gaps.

Last weekend, I tested a bunch of tools/models on implementing a design (for an ads storefront for Ben’s Bites) from a screenshot. I found:

• Claude Design > Magicpath AI > Raw models (like Gemini 3.1 Pro/Opus 4.6 in their web apps), when it comes to understanding the concept and making something usable, not just copying the pixel-by-pixel look (ironically, Gemini won that).

• When asked to turn designs into a real working app, there was a major drift in how the apps looked. Opus 4.7 did better than GPT-5.4 at visually matching the reference screenshot. Though GPT-5.4’s code was more functional, and the unseen pages (like the admin panel) had a consistent design with the rest of the app.

Also, in many cases, the assets (hero image, icons, background textures) make the UI in a “generated image” stand out. When replicating that UI from a screenshot, you get the barebones UI with the correct buttons and the layout, but without those assets, and the output falls short of expectations.

Ben’s Bites is brought to you by TinyFish

Funny how AI agents can write entire apps but can’t work on the live web. Playwright scripts break, raw fetches eat your context, bot detection blocks you, nothing’s scalable. TinyFish gives search, fetch, stealth browser, web agent, all managed in one API. Try it free. Comes with a CLI + Skill.

Headlines

• OpenAI has a new product for Business, Enterprise and Edu users - Workspace Agents. Codex-powered agents inside ChatGPT with a persona, task and access to external tools (like Linear) and accessible for Slack as well. These agents will also replace custom GPTs down the line (finally). Read more.

• Gemini Deep Research API now offers two configurations based on 3.1 Pro. It claims the best performance in web research and finding hard facts. Plus, it gets MCP support and can create charts using Nano Banana or HTML.

• Cursor and SpaceX are working together - Cursor will train coding models on SpaceX’s GPUs and likely share them with xAI. SpaceX can, in turn, acquire Cursor later this year for $60B, or pay $10B for the partnership if it doesn’t. On a similar note, Thinking Machines also just signed a multi-billion-dollar Google Cloud deal.

• Give your Droid a computer - You can now give your Droid an always-on machine with its own filesystem, credentials, and config for it to keep working on your tasks. This can be in the cloud (managed by Factory), or you can bring your own device.

Subscribe now

My feed

• Chronicle - Cursor for slides. Never build a deck from scratch again. Turn ideas into stunning presentations in minutes.*

• ChatGPT for Excel and Google Sheets are now in beta - build new sheets, fix formulas, explain models, and update workbooks in place. (read more)

• /ultrareview in Claude Code (research preview) lets you run bug-hunting agents in the cloud before merging riskier changes like auth, data migrations, or other critical code paths.

• OpenAI built an open-source viewer for chat data and Codex session logs - Euphony.

• Sierra is piloting an AI-native interview - debugging/review focused interviews where candidates improve a medium-sized codebase with coding agents.

• ml-intern from Hugging Face - open-source research agent to come up with experiments, and run them.

• Clawputer - Managed OpenClaw agent inside an always-on sandbox.

• Kami - design skill for AI-native docs, resumes, portfolios, long docs, and slides.

• noscroll - an AI that doomscrolls X for you and texts you just the signal. In my experience, this is easy to claim and hard to get right.

• Monologue has a new Notes feature for thinking out loud when you don’t know the exact words you want to dictate.

• Fin is moving beyond customer support into sales - using the same business context and integrations to qualify leads and book meetings.

• Perplexity post-trained a Qwen-based model to handle search and tool calls for cheaper, and it’s already serving a meaningful chunk of traffic.

• The next Slack won’t look like Slack, and Ando looks like one early attempt at that.

• Frontend in 2026 - for and against the frameworks and abstractions dominant today.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’ve been playing a lot with Claude Cowork for my talk at Stanford later today. It’s comically bad for the average jane.

A lot of capabilities are enabled using connectors and plugins, but if you don’t know that, good luck getting it to do anything. Can’t seem to send emails, install a skill or tell me about what potential connectors are there.

Scheduled tasks in Cowork stop when you shut the lid, but Routines (similar thing) in Claude Code do not. Cowork is just now getting Artifacts - the mini apps on the Claude chat app that started the vibe-coding wave.

I can search for all of this, yes. but an average user will not. and they’ll walk away thinking AI is hype for the next 6 months/a year.

Anyway, the Cerebral Valley AI Summit is coming back to London on June 24th. I’m planning to be there.

Ben’s Bites is brought to you by Gauntlet

Most RAG systems fail in production. Gauntlet Night School on Wednesday, April 22 covers how to build one that doesn't — live and free - Register. Or go deeper: become an AI-native engineer at no cost. Cohort starts April 27. Apply now

Headlines

• Opus 4.7 is out - much better at vision (interpreting images) and efficient at using reasoning tokens. A new xhigh level of thinking now sits between “high” and “max”. I’ve been using it over the weekend at xhigh and didn’t face any issues (despite the general sentiment on Twitter that 4.7 is a regression).

• Claude also has a Design tab now - a canvas-like interface with chat on the sidebar to explore wireframes or create high fidelity prototypes. It asks you 5-10 questions via an interactive form and then gets building. I found the image → design workflow to be really good in the prototype mode. Has separate limits while in research preview, but expect the weekly limits to only last for 2-3 big generations max (on the $20 plan). Check out Peter’s demo across multiple use cases.

• Codex got a few updates: 1. Computer Use - it can now use apps on your Mac. I’m not completely sold on it and have some questions, but it does seem to work a lot faster than previous iterations of computer/browser use demos. Also, it works in the background, so your Mac is free for you to use. 2. Chronicle, an opt-in preview that uses recent screen context to build memories. 3. A bunch of plugins, including image generation, so that Codex can be the superapp to use AI.

• Factory AI is now valued at $1.5B after their latest raise of $150M. Try the new desktop app for Droid, Factory’s coding agent, with 50% off on Opus 4.7 till 30th April.

Subscribe now

My feed

• Your tools are full of data, but you’re still guessing. Backstory - Focus on what drives revenue, based on how your business actually sells.*

• Google AI Pro and Ultra subs now work with AI Studio - higher rate limits for vibe coding in the playground.

• Julius can now generate slide decks with charts and tables, exportable as pptx.

• Galaxy Brain - an operating system powered by local files. I’m an investor.

• Kimi 2.6 Code - a Claude Code-style terminal built specifically for Kimi K2.6.

• Moondream Lens - fine-tune a vision model to production accuracy in hours with as few as 20 images.

• Zapier’s AutomationBench measures real work done by models - CRM updates, inbox follow-ups, multi-step tool chains. No model has cracked 10%.

• How AI made me a builder without losing my taste as a designer.

• Quiver upgrades its models for vector generation - Arrow 1.1 and Arrow 1.1 Max.

• acceptmarkdown.com - checks whether your site returns Markdown correctly for agents.

• The 5 stages of AI grief.

• HeyGen open-sourced HyperFrames (HTML to MP4).

• Vercel was breached via an employee’s account on another AI product. Vercel says affected customers have already been contacted. Quick check to make sure you’re safe.

• Three OpenAI leaders left the company - Kevin Weil (CPO, then OpenAI for Science lead), Bill Peebles (Sora co-creator), and Srinivas Narayanan (CTO for B2B Applications).

• Skill drops: skill-creator, make-interfaces-feel-better, and gpt-taste for Codex.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Boarding my flight to SF very shortly, and I got an email to let me know - no WiFi today. Uh oh. I was kinda hoping my 11 hours uninterrupted hours without the kids would be productive for once (I’m usually a very OOO long-hauler, no internet). But I still have some work to polish this talk I’m giving on Tuesday.

I’m also in town looking to deploy $100k cheques to dev tools and infra founders, plus see some of my wonderful LPs and meeting new ones. Ben’s Bites Fund II has already started investing.

So my flight… I’ve had to hurriedly download a few local models so I can use my agents offline and I think, so far, Gemma 4: 26b is going to be my choice.

We’re so spoiled today with fast intelligence at our fingertips and it’s funny how used to the new intelligence levels we get

Local models are slow to boot up (you’ve got to be more mindful of what context is being loaded on startup (so I’m running with no-skills to get it to go faster, I can call the skills when I want — maybe I’d actually prefer to do that generally 🤔). And they feel pretty slow to do work, but only because of said spoils.

I’ve been in the weeds of context management recently because of the course I’m working on. And it’s been useful to just remind myself about how prickly it can be;

• If an agent runs web searches - presumably you didn’t read them, its gobbling up context from content you do not know is 1. right, 2. not ai-slop, and 3. by a source you’d recommend.

• Little (or big) lines of slop, misdirection, misinformation slip in to the context and compound over time

• Reaching ~60% of a context window is probably the limit of where you want to be

• Use other sessions as context-gathering sessions, if there’s lots of documents then create one summary file with the information (and try to read or at least skim it! - I am trying, promise)

• I don’t trust 1M context windows, there’s a great post by Thariq from Anthropic below about this window. I shouldn’t need my context for my tasks to need perfect recall beyond ~150k tokens, that’s a lot of words. Only until 1M context windows are the norm, the models dont forget anything and help clean polluted context along the way!

Anyway, got to head to the gate! This was a little different of an intro, let me know if you liked it. I need to share more as I’m learning (or diving deeper).

Ben’s Bites is brought to you by Attio, the AI CRM

Honestly, no one gets excited about a CRM. But then they try Attio. It connects to Claude Code and n8n through its MCP server, completely bridging the gap between my customer data and apps. Wait, there's more, like flagging churn risk and turning customer feedback into Linear projects. Try it now.

Headlines

• Claude Code’s desktop got a redesign. Brings many CLI-only features and more (like split windows for multiple sessions) to the desktop app. Big improvement, but still a lot is missing. It picks up some CLI sessions but not all, opening/editing files isn’t obvious, and it keeps asking for permission even with “bypass” settings on.

  • Gemini also has a native Mac app now. But it’s light on features - no Gems, no notebooks - and the design feels rough to say the least.

• New models - GPT-5.4-Cyber from OpenAI, fine-tuned for cybersecurity, with limited access to trusted partners. And Gemini 3.1 Flash TTS from Google - better voices, audio tags for controlling tone and pacing, and 70 languages.

• Routines in Claude Code are now in research preview - set up a prompt, a repo, and your connectors once, then run it on a schedule (or via API/GitHub trigger). Runs on Anthropic’s infra, so you don’t need your laptop open. Basically, extended cron jobs. OpenClaw calls these heartbeats.

• With the latest update to OpenAI’s Agents SDK, you can run Codex-style agents in production without building the whole harness yourself. You get sandboxed execution, computer-use, skills, memory, and compaction built in.

• Most RAG systems return wrong answers with complete confidence. Gauntlet's free Night School covers how production AI engineers actually fix that — setup, evaluation, the full loop. Wednesday, April 22. Register free*

Subscribe now

My feed

• Skills in Chrome let you save prompts as reusable one-click workflows that run on whatever page you’re viewing.

• Cursor can now respond with interactive canvases - dashboards and custom interfaces instead of just text.

• Resend shipped a new email editor with BYOA (bring your own agent). There’s a built-in LLM, but you can also MCP into the editor with your own setup.

• Sparkle v4 from Every - let AI organise your filesystem like you would.

• Daniel pointed an agent at 5 years of home-building emails (511 events, 690 documents, 170 finance records) and got back a full project timeline in ~$500 of Opus tokens.

• Impeccable v2 - the design skill for coding agents. v2 adds a CLI scanner (works without an LLM), a Chrome extension, and a /shape command that runs a design interview before writing any code.

• Using Claude Code - guide on session management, compaction, and the 1M context window.

• 30 min tutorial on building software with agents in Cursor.

• Lindy AI’s founder says GLM 5.1 will likely become their default over closed-source models for most use cases, saving them a bunch on inference (their biggest cost, more than payroll).

• OpenRouter now offers video generation models with one universal API across all video models.

• Copilot in Word now tracks changes and leaves comments.

• Windsurf 2.0 - Manage all your agents from one place and delegate work to the cloud with Devin.

• Gradient Bang - a fun multiplayer game with subagents in space. Built with Pipecat, Supabase, and open-source.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

I’ll be heading to SF on Thursday — speaking at Stanford the week after. I’ll be delivering live the ‘how to use agents if you’re not technical’ course I’ve been working on (ahem, still working on).

I’ll be in town chatting to LPs (my 2 prev funds; 2.3x MOIC, 43% IRR (2023), and 4x MOIC, 28% IRR (2020)) and founders (I invest in AI dev tools and infra!) - let me know who I should try and meet or any cool things going on.

The latest leak from Claude-land shows Anthropic is adding lovable-like features for building full-stack apps right inside Claude. Codex apparently had a similar leak (since deleted).

There’s a new term going around for products that agents can use - headless SaaS. Box’s CEO says that “Enterprises will kick out vendors that don’t make it easy/economical for agents to use their product”. And you can’t just wrap your APIs into MCP/CLIs and call it a day.

Talks at AI Engineer were across the spectrum from “code is a liability” to “slow the fuck down”. Alex has a good recap.

Ben’s Bites is brought to you by Cloudera

AI is booming, but ROI is lagging. The problem isn’t AI—it’s data readiness. Most enterprises aren’t as prepared as they think. See where things really stand in Cloudera’s Data Readiness Index.

Headlines

• Claude Cowork is now generally available - out of research preview after 12 weeks with millions of users. Also, Claude for Word is in beta - draft, edit, and revise documents from the Word sidebar with edits showing up as tracked changes. (Team and Enterprise plans)

• New Claude Code features - /ultraplan lets you build and edit a plan on the web, then run it in your terminal. The Monitor tool lets Claude watch for events in the background instead of constantly checking - saves a lot of tokens. Plus, an advisor strategy on the Claude dev platform that pairs Opus with Sonnet for better performance at similar/cheaper costs.

• OpenAI added a new $100 plan. Their own wording is too complex, so I’m stealing Theo’s. Basically, $100 plan = 5x compute, $200 = 10x compute (where $20 plan = 1x). The bigger plans have 2x boost till May 31st, so effectively, $100 = 10x and $200 = 20x compute.

• LlamaParse was built to tackle your most complex PDFs with the highest accuracy & the lowest cost. Now we want to put that to the test. Submit your ugliest PDF (dense contracts, financial tables, etc.), and judge the output vs your current OCR to win a Mac Mini. Can your doc outsmart LlamaParse?*

Subscribe now

My feed

• Anyone can make a game now. Yoroll turns a rough idea into a branching, playable video-native game. Join waitlist to try.*

• Cursor cloud agents can now attach screenshots and demos to PRs they open. Your team reviews the artifacts directly in GitHub.

• Shopify AI Toolkit - manage your Shopify store with your favourite agent. Works with Claude Code, Codex, Cursor, VS Code, and more. (docs)

• Open Agents by Vercel - open-source template for building coding agents.

• How Missions work - from the Factory team.

• Agentic parenting and TradClaw.

• The anatomy of an agent harness and a tutorial/walkthrough to build your own harness.

• How I became technical AF.

• Personal wiki tools inspired by Karpathy’s knowledge base idea - LLMwiki, Wiki OS, Hatch, and GBrain.

• Evo, another Karpathy-inspired CC plugin, optimises your code through experiments.

• Ramp says their entire company is AI-pilled. They even built a new internal product called Glass to give every employee an AI coworker.

• getdesign.md - browse design systems of popular applications as markdown files. Preview instantly, install with one command.

• Aqua Voice - premium voice keyboard for every app on your phone. Now on iOS.

• Gitinspect - replace “hub” with “inspect” on any GitHub URL to chat with the repo. Runs in the browser, everything stays local.

• Cloudflare Sandboxes are generally available - comes with a terminal, interpreter, live preview URLs & secure credentials. Sleeps when idle and wakes on demand.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail sourced from Twitter

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

A mis-timed blog last week leaked Anthropic’s next model - Claude Mythos. Well, it is real and has massive improvements on benchmarks over Opus 4.6:

• 53.4% → 77.8% on SWE-bench Pro

• 65.4% → 82% on Terminal-Bench 2.0

but we are not getting access to it anytime soon. Why? because it is really good at finding and exploiting software vulnerabilities. On Firefox exploit generation, Opus managed 2 working exploits out of hundreds of attempts. Mythos hit 181.

It found many-decades-old bugs in critical software projects like OpenBSD (27-year-old bug), FFmpeg (16-year-old bug) and more.

Instead of releasing it publicly, Anthropic is giving 12 companies access to a preview version of Mythos under “Project Glasswing” to find vulnerabilities in critical software. Anthropic is committing $100M in model usage credits and $4M in donations to open-source security orgs under this project.

Theo made a video on this, and I like his point: “Mythos is to Opus what Opus is to Sonnet.”

I tweeted a list of companies that Meta has acquired in the past year without anything to show for it, and soon after, Meta released details about their latest model - Muse Spark. At a glance, it sits somewhere between Sonnet 4.6 and Opus 4.6. Not usable yet: API access is coming, and there are promises about open-source too (rip llama).

Many people are dunking on Meta for its not-so-frontier model release after spending billions and a year of silence, but I think it’s a good step ahead. Plus, have you used Instagram search over the past couple of months? It’s gotten really good courtesy of AI.

As always, good recap from Ethan Mollick on the state of frontier models: Google, OpenAI and Anthropic lead, Meta joins the pack for now while xAI has fallen off, and the best Chinese models are still 7-9 months behind.

ps: Factory’s desktop app is now out of beta. It comes with a cloud computer, the ability to use other apps on your device, and, of course, the ability to run and manage multiple Droid sessions easily.

Ben’s Bites is brought to you by Attio, the AI CRM

Honestly, no one gets excited about a CRM. But then they try Attio. It connects to Claude Code and n8n through its MCP server, completely bridging the gap between my customer data and apps. Wait, there's more, like flagging churn risk and turning customer feedback into Linear projects. Try it now.

Headlines

• Claude Managed Agents - You can use Claude’s developer console to build and deploy agents and let anthropic handle the infra for it, vs building it yourself. For example, Notion is using managed agents to build a “delegate tasks to Claude” feature. (Anthropic’s engineering blog on building this).

• Cursor has a new design mode to annotate and target UI elements in the browser. Plus, run Cursor on any machine and control it from anywhere, including your phone.

• Gemini app finally has projects - they call it notebooks. Similar features as Claude/ChatGPT projects - move chats in/out of notebooks, notebook-specific files and memories, with the additional feature to sync these notebooks between the Gemini app and NotebookLM.

• Clicky is an ambient AI buddy on your Mac. It sees your screen, talks to you and points at things to guide you (demo). Farza built (and open-sourced) it as a learning tool, but people are using it for everything.

• Choosing an accurate speech-to-text model is harder than it looks. Benchmarking one is even harder. See why standard word error rate falls short, and what better STT evaluation actually looks like.*

Subscribe now

My feed

• Chronicle: Cursor for slides. Never build a deck from scratch again. Turn ideas into stunning presentations in minutes.*

• OpenRouter Spawn - Deploy OpenClaw and other agents to the cloud of your choice. Works with all models on OpenRouter.

• Zapier’s SDK is now open to everyone. Programmatic access to all of Zapier’s capabilities. Free to use in beta. (docs)

• Kiro.dev (spec-driven IDE from Amazon) is bringing its startup credits program back for startups with up to 30 people.

• Cogito - Markdown editor for Mac. I’ve been using Clearly (recently updated) for the last few weeks to simply view and edit md files.

• Graphify - Turn any codebase or folder into a queryable knowledge graph.

• Pi and Mario (the maker of Pi) are joining Earendil, the company by the creator of Flask. The core harness stays open-source. New features will be a mix of enterprise & fair source (proprietary now, open-source later).

• Impeccable - Free design skills for coding agents with 21 commands to audit and fix common mistakes.

• Superset and Builder 2.0 - two new UIs for running parallel agents. Superset is more like Codex (terminal-first, worktrees), Builder is more kanban-style with Slack/Jira integration.

• CSS Studio by Motion - Make design changes by hand on your website in the browser, then pass them over to your agent for implementation.

• S3 Files from AWS allows storing data as a file system, making it easier for agents to use.

• Every is running two parallel org charts - one for humans and another for each employee’s openclaw agents.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

But OpenClaw founder, Peter, is a (good-kind of) maniac who’s trying to get gpt5.4 working like opus in the tool (since OpenAI acquired him).

Unfortunately it’s just left the whole community confused as to where and when they can use their Claude Code subscriptions outside of the cc harness itself. Myself included.

There’s a new take on knowledge bases that I think is very interesting by Andrej Karpathy. Memory and file organisation for retrieval got a big focus with the OpenClaw hype but this approach groups things topically and then enhances with summaries, backlinks and wikis.

Related, Farza built Farzapedia, a personal Wikipedia generated from 2,500 diary entries, notes and messages. Built for his agents to crawl whenever needed. Karpathy’s take on Farzapedia

OpenAI acquired TBPN -- good pod to get a peek into the heads of the hosts. A lot of people on X are trying to sound smart with ‘this is why they acquired them’ but I’m more in the Ben Thompson camp… I don’t really know why either party needs each other here? TBPN is loved, growing, and making a small fortune with their (also loved!) ad business. Why does OpenAI need them? Other than the fact that they don’t like traditional media and could use TBPN as their main channel to people that matter.

Ben’s Bites is brought to you by Reevo

The best sellers aren’t the ones with the most sophisticated tech stack. They’re the ones who spend the most time selling. Reevo is the AI-native revenue intelligence system and CRM that every team deserves. One Platform. From prospect to close in a single tab. Go Stackless. Reevo.ai

Headlines

• Google released Gemma 4. 4 new open-weights models - two for powerful laptops and desktops (26B MoE and 31B dense), and two tiny ones for mobile phones (2B & 4B). Relevant for companies that want to fine-tune them, not for direct use.

• Cursor-3 adds a new standalone window for running agents. Has some neat additions for local to cloud handoff, working across different projects.

• Anthropic’s run-rate revenue hit $30B, up from $9B at the end of 2025. Lenny’s podcast with Anthropic’s head of growth is worth a listen — they added $6B ARR in February alone.

• Adapt is your company’s AI computer. It connects to your tools and team, autonomously learns your business, and gets smarter with every task. Pull live data, run scheduled workflows, and deploy internal apps everyone can use, all from Slack.*

Subscribe now

My feed

• Salmex I/O — Won't forget you. Gets things done. Answers to no one. And is actually safe! What a local AI agent should be.*

• What could “devrel” mean in a world with slot machines spitting out code and opinions?

• agent-browser by Vercel can now also work as a browser agent with a chat-bar on the side.

• Turning any documentation website into a filesystem for agents to navigate with tree, grep, cat, etc.

• The CEO of Brex claims he is running the $5B+ company with OpenClaw.

• Growth ideas that helped Michelle scale Warp to 500k+ weekly active developers.

• The era of web apps is over - Bret Taylor

• Deep Extract adds an agent in the loop for extracting data out of complex documents that verifies and corrects itself.

• Samuel is building skills (like this AI SDR one) for the Replit Agent and sharing them on X. Cool way to market agents to non-technical users.

• Simon is the most up-to-date person on LLMs. He was on Lenny’s pod to talk about the changes in AI coding over the last few months.

• Intercom is making its agent Fin available via API for companies with contracts starting at $250k per year. (read more)

• Loophole translates your natural language moral beliefs into codified laws and then tries to break them.

• Ex-OpenAI board member (left after the Sam Altman firing fiasco) argues for retiring the term AGI in favour of specific milestones.

• You can now use your agent to apply for a job at Every.

• Autoagent - An agent to help optimise agents.

• Planning is dead

• A YC partner one-shotted a tool to manage his inbox with voice. Another YC company launched an AI secretary that manages your email like you.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com

Re unreleased/hidden features (because I know you wanna know); “Proactive” mode (AI works autonomously without prompts), frustration/anger detection via keyword patterns is marked ‘negative’ in their analytics, a Tamagotchi-like “/buddy” companion, background daemons (like OpenClaw uses), undercover commit-hiding mode, and more feature flags/roadmap items.

The community ran rampant cloning the codebase on GitHub (which Anthropic since sent DMCA notices to get them taken down). But some developers ported the code to other languages (Python and then Rust) - which has copyright greyness and if Anthropic push to get them removed, may bring up questions on their own copyright issues.

Since it was April Fools yesterday, many launches are indistinguishable from real ones - I think it’s getting harder to spot in the AI age too.

But some folks use it as an opportunity to launch things, like Gumroad replacing their CEO with an AI Agent. Sahil (the founder) is a friend and previously invested in my last company. He’s super smart and I don’t think this is a prank at all. He’s the kind of person to try wacky, out-there things to see if they stick.

Some companies actually do April Fools well, like ElevenLabs. Last year, they made dogs talk, and this year, they are partnering with the 3000-year-old ElvenLabs.

Ben’s Bites is brought to you by LlamaIndex

Your agent can write code but it can’t accurately read PDFs. Classic.

LlamaParse 🦙 reads docs with 99%+ accuracy on complex PDFs–like SEC reports, invoices, or research papers. Give your agents the context they need.

Sign up today to get 20,000 free credits with coupon: BENSBITES20 (ends 4/10)

I invested in LlamaIndex when scouting for a16z a few years ago

Headlines

• Claude Code now renders a full UI in your terminal to solve the flickering. It shows “N new messages” when you scroll up, gives you full mouse support, and constant memory use. It’s experimental for now; upgrade to the latest version and run using “CLAUDE_CODE_NO_FLICKER=1 claude”.

• OpenAI closed its $122B raise at a $852B post-money valuation. It’s making about $2B a month in revenue, 40% of which comes from the enterprise. The blog post outlines a lot of their plans that were previously just rumours/leaks from various newsrooms.

• Slack is turning Slackbot into a desktop companion. Select anything on your screen, pass it to Slackbot, and it can answer based on the context from your Slack workspace. It has 30+ new capabilities, including meeting transcription, reusable AI skills, MCP client and a native CRM.

• Softr AI builds the tool your business needs, not a prototype. It wires the whole thing: database, logic, logins, permissions, & security (even hosting) with a visual editor to fine-tune. Think client portals, internal tools, CRMs, and more. No code required. 👉 Start today with 200 free AI credits.*

Subscribe now

My feed

• Attio - The AI CRM that does the work for you.*

• AgentOS by Rivet.dev - open-source operating system for agents. ~6 ms coldstarts, 32x cheaper than sandboxes.

• Mario, the founder of Pi, asked around for something like Google Docs, but for markdown. He then vibed Jot in an hour.

• Your agents can now access Supabase’s full documentation as a virtual filesystem. I think Mintlify could offer this to their customers, similar to how they added the “Copy as Markdown” button to the docs hosted via Mintlify.

• Vibe coding SwiftUI apps is a lot of fun.

• Simple Mac app to run skills on a schedule via Claude Code.

• Exa’s new product, Monitors, returns fresh results from the web for your query on schedule. Feels relevant for building agents.

  • Yutori’s Scouts do something similar for humans. They also launched Yutori’s desktop app for local-only access to logged-in websites when creating these “what’s new for your query” reports.

• The case against personal software.

• Cloudflare launched a successor to WordPress with a focus on plugin security.

• How Zapier measures the AI fluency of every hire across all departments.

• Block wants to move away from traditional human hierarchies and transition to one guided by an intelligence layer.

• Microsoft is working towards adding OpenClaw-style personal agents to M365.

Afters

Share Ben's Bites

• Find me on X, Linkedin, or YouTube

• Read about me and Ben’s Bites

• 📷 thumbnail by @keshavatearth

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com or k@bensbites.com