This week something clicked. I stopped thinking of AI as a tool I use and started thinking of it as a colleague I delegate to. That shift changes everything about how you build.
π₯ The Big One
Anthropic dropped 9 features in 7 days. Claude isn't a chatbot anymore.
Let me walk you through what just happened, because I think most people are sleeping on this.
Anthropic launched Claude Computer Use, which lets Claude directly control your desktop. Mouse, keyboard, browser, screen. You tell it what you want done, and it does it. Not in a sandbox. On your actual machine. This shipped alongside Claude Cowork Projects, which gives Claude persistent project-based workspaces with local files, memory, and scoped context. That means Claude finally remembers what you're working on between sessions.
Then they dropped Claude Code Scheduled Tasks. A `/loop` command with natural language cron. Tell Claude to check your CI every hour, monitor logs, open PRs when tests pass. It just... runs. Locally or in the cloud.
Put these three together and you don't have a chatbot. You have a coworker that sees your screen, knows your codebase, and works while you sleep. 9 features in one week. Anthropic isn't iterating. They're redefining what an AI product looks like.
If you're still copy-pasting code from ChatGPT into your editor, you're already two generations behind.
"Claude Cowork represents our vision for AI as a true collaborator, not just a responder." β Anthropic
π οΈ What I built this week
Anthropic shipping a coworker-level AI lit a fire under me. Here's what I've been building.
-
OpenClaw agent orchestration β Rebuilt how my AI agents coordinate tasks across the PAPAFAM division. Claude Code Scheduled Tasks inspired me to rethink background automation entirely. Agents now self-delegate overnight work without me touching a thing.
-
Newsletter pipeline with Scout + Ink β This newsletter you're reading was researched by one agent, written by another, and assembled with zero manual copy-paste. The AI coding stack is eating its own dogfood.
-
Agentic content workflow β Integrated Vercel's new AI coding agent plugin into our deployment pipeline. Claude Code now has full context on every Vercel project. Deployments went from 4 manual steps to zero.
β‘ What shipped this week
1. Cursor drops their own coding model, and it beats Claude Opus
This is the story nobody expected. Cursor, the IDE company, trained their own proprietary model called Composer 2. It scores 73.7% on SWE-bench Multilingual and 61.3 on CursorBench. For context, that puts it ahead of Claude Opus 4.6 on real-world coding tasks. And the pricing? $0.50 per million input tokens versus Opus at $15. That's 30x cheaper for comparable performance.
Cursor went from "nice wrapper on top of Claude" to "we're competing with Anthropic on the model layer." That's a massive strategic shift. If you're building with AI, this changes the math on which model you pick.
"Feels like Claude 4.6 Opus but faster and more cost-efficient." β Cursor Community
2. OpenAI acquires Astral, and Python developers should care
OpenAI just bought Astral, the company behind `uv`, `Ruff`, and `ty`. If you write Python, you've used their tools. 126 million downloads per month. And now they're part of the Codex division.
This isn't about code generation anymore. OpenAI wants Codex to own the entire developer workflow: package management, linting, type checking, and code generation all under one roof. That's a fundamentally different product than "paste a prompt, get code back."
The open source community has mixed feelings. But strategically? This is OpenAI signaling they want Codex to be the IDE, not just the model inside one.
"The goal is to move Codex beyond code generation toward full dev workflow participant." β Ars Technica
3. Google Stitch 2.0 brings AI-native design to your workflow
Google launched Stitch 2.0, an AI-native infinite canvas where you design UIs with natural language and voice commands. But here's the part that caught my attention: the DESIGN.md system. It's an agent-friendly markdown file that exports and imports design rules between tools. Your AI coding agent can read it and understand your design system.
It ships with an MCP server, an SDK, and exports directly to Figma. 640 upvotes on Product Hunt, the #1 launch of the week. Google is quietly building the bridge between "AI generates code" and "AI generates code that actually looks good."
"The new design agent tracks progress across multiple ideas in parallel." β Google Stitch
4. Google AI Studio is now a full-stack app builder
This one flew under the radar. Google AI Studio upgraded from a model playground to a full-stack app builder. Describe what you want in plain English and it generates a working app with auth, database, and API connections. Firebase integration is baked in.
We've now got Bolt, v0, Lovable, Replit, and Google all competing in the "describe an app, get an app" space. The difference? Google has Firebase, Gemini, and the infrastructure to actually scale what gets built. This isn't a toy demo. This is Google saying "we own the full stack from prompt to production."
"Build full-stack applications with AI Studio and Firebase, from description to deployment." β Firebase Blog
5. MiniMax M2.7: the model that helped build itself
This is genuinely wild. MiniMax's M2.7 model contributed to 30-50% of its own reinforcement learning research. A model helping to train the next version of itself. It scores 56.22% on SWE-Pro, which matches Codex. The hallucination rate dropped to 34% compared to Claude Sonnet 4.6's 46%.
And the pricing: $0.30 per million input tokens, $1.20 per million output. For a model that matches top-tier coding benchmarks, that's absurdly cheap. Keep this one on your radar.
"M2.7 contributed to 30-50% of its own RL research, marking a step toward self-improving AI systems." β MiniMax
π§° Worth your time
-
Bench for Claude Code β Free session observability for Claude Code. It records every tool call, file change, and reasoning trace automatically. Step-by-step replay and searchable sessions you can attach to PRs. Built by ex-Google and Meta engineers. If you use Claude Code daily, this is a no-brainer.
-
Tobira.ai β "LinkedIn but your AI agent does the networking." An open network where AI agents discover deals and negotiate on behalf of their humans. Agents get public identities and there's a human approval gate before any contact exchange. 496 upvotes on Product Hunt. The agent-to-agent economy is coming faster than expected.
-
Working at Machine Speed β A thoughtful deep dive on how AI coding is reshaping dev roles. Martin Fowler's take: developers are moving "on the loop" not "in the loop." You're becoming an agent manager whether you planned to or not. Worth reading with your morning coffee.
My weekly message to YOU!
Every tool I mentioned this week has a free tier. Every single one. Claude Code, Cursor, Stitch, AI Studio, Bench. The barrier to entry is literally zero.
So here's your challenge: pick one tool from this newsletter that you haven't tried, and build something with it before next Tuesday. Not a tutorial project. Not a todo app. Something you'd actually use. A tool for your workflow. An automation for your team. Something real.
The developers who compound skill with AI tools this year will be unreachable by the ones who wait. That gap grows every week.
What's the one tool you're going to try? Reply and tell me.
I read every single one.
Talk soon PAPAFAM,
Sonny ππΌ
ππ½Don't forget to follow me across socials!
Responses