- Building AI Agents
- Posts
- Anthropic Buys Into Computer Use
Anthropic Buys Into Computer Use
Plus: Codex ships on Windows, Gemini's Flash-Lite price war, Europe's first agent payment, open-source sandboxing drops, and more...
AP Patel
March 05, 2026
In partnership with
Edition 163 | March 5, 2026
Even AI agents are out-earning junior devs now.
Welcome back to Building AI Agents, your guide to everything new in the field of agentic AI!
In today’s issue…
OpenAI Codex lands on Windows
Anthropic acquires Vercept for computer use
LLMs can unmask anonymous users at scale
Europe's first AI agent payment goes live
Why your team is your most expensive integration tool
…and more
🔥 IN CASE YOU MISSED IT
Readers’ favorite items from the past week
📌 THE THURSDAY BRIEFING
OpenAI
📡 Signal: OpenAI launched the Codex app for Windows with native sandboxing and PowerShell support. The Mac version crossed 1 million downloads in its first week, and over 500,000 Windows developers were on the waitlist. ChatGPT Free and Go users can now try Codex for the first time.
🤖 For builders: OpenAI worked with Microsoft to build native Windows sandboxing from scratch because the OS lacked the isolation primitives that macOS offers out of the box. The result is the only major coding agent with a first-class Windows experience. Nearly 50% of professional developers use Windows (Stack Overflow 2025), and until now they were stuck on CLI or WSL workarounds. With 1.6 million weekly active users already on Codex, OpenAI is making the case that your coding agent should live in a dedicated app, not an IDE sidebar. Cursor, Windsurf, and Claude Code are all watching.
📡 Signal: Anthropic announced it acquired Vercept, a Seattle startup that built vision-based agents capable of remotely operating a MacBook. The Vercept team, including co-founders from the Allen Institute for AI, will join Anthropic to work on Claude's computer use capabilities. Claude Sonnet 4.6 now scores 72.5% on OSWorld, up from under 15% in late 2024.
🤖 For builders: This is Anthropic's second acquisition in three months (after Bun in December), and both are aimed at the same thing: making Claude the best agent for doing real work on real computers. Vercept's product Vy could see screens, understand UI elements, and take actions the way a human would. That capability is now folding directly into Claude. Meanwhile, UiPath's stock dropped on the news, signaling that Wall Street sees computer use agents as a direct threat to the $13B RPA market.
🤝 WITH ATTIO
Attio is the AI CRM for modern teams.
Connect your email and calendar, and Attio instantly builds your CRM. Every contact, every company, every conversation, all organized in one place.
Then Ask Attio anything:
Prep for meetings in seconds with full context from across your business
Know what’s happening across your entire pipeline instantly
Spot deals going sideways before they do
No more digging and no more data entry. Just answers.
🤖 AGENT OF THE WEEK
👋 Welcome back to Agent of the Week!
If you've ever done a strategy call with a client, you know the drill. You hop on a call, walk through their current processes, take a bunch of notes, and then... spend the next 2-3 hours turning those messy notes into something presentable.
Flowcharts, summaries, pain points, tools they mentioned. All manually built in Google Slides or Canva, slide by slide.
It's important work because it's what you send back to the client to say "here's what we found." But it's also the kind of work that makes you question your life choices at 11pm on a Wednesday.
So I built an agent that does it for me using Claude Cowork.
Meet the Process Map Generator Agent.
Here's the setup: I had a strategy call where we walked through five different business processes, things like client onboarding, weekly reporting, campaign launches, lead follow-up, and invoicing. Just a normal discovery call where you're mapping out how the business actually runs day to day.
Normally after that call, I'd spend hours rewatching the recording, pulling out each process, sketching out the steps, identifying where things break down, and building out branded slides to send back.
Instead, I dropped the call transcript into a Cowork project with a custom skill I built, and let it do the work for me.
Five processes extracted. Five flowcharts generated. Pain points identified. Tools documented. All formatted into a branded PDF report with my company colors, logo placement, and consistent slide layout.
The whole thing took about 5 minutes. Something you can let run while you read emails.
🧠 How It Works
Record Your Call → Get a Transcript: Fathom, Fireflies, Otter, Zoom's built-in transcription, whatever you use. Just need the raw transcript.
Build the Skill → One-Time Setup: Three files in a Cowork project: a SKILL.md with instructions, a Python script for PDF generation, and a brand config with your colors and company name. Set it up once with ChatGPT or Claude, reuse forever.
Drop Transcript → Extract Processes: Claude reads the full conversation, identifies every process discussed, and structures each one with steps, decision points, pain points, and tools used.
Flowchart Generation → Branded PDF: The script builds a PDF deck automatically. Title slide, then one slide per process with the flowchart, summary, pain points, and tool tags all laid out.
Review and Send → Done: Quick once-over and modifications I can ask Claude to do. Professional process mapping report delivered in the time it takes to make coffee.
From a raw call recording to a branded deliverable, with zero manual diagramming and zero copy-pasting into slide templates. Great to show the client and great for our internal team to reference.
This is what I mean when I talk about workforce multiplication. The strategy call itself is the valuable part, that's where the human insight happens. But the documentation? That's just formatting what was already said. Let the agent handle it.
And if you're interested, this is just one of the many multi-step agent workflows you can learn to build inside our Building AI Agents Community.
Till next week,
✌️ AP