Welcome back. The era of the trillion-dollar AI IPO is officially upon us. Last week Anthropic filed a confidential S-1 to go public, and now OpenAI has followed suit. Oh, and there’s also the small matter of the SpaceX IPO, which is looking to go public at a $1.8 trillion valuation.

Also: How to make Claude Code review its own work, a skill that catches AI's bugs before they ship, and why the "Godmother of AI" says AI must change how we teach and evaluate students.

Google turns NotebookLM into a code-running research agent: The search giant just handed its research tool a secure cloud computer. It can now write code, use over 100 software skills, and pull fresh sources straight from the web. Engineering teams can turn dense specs into guides, decks, and roadmaps, then export the results as charts or spreadsheets. Google says it now runs on Gemini 3.5 and won 65% of evals against the old version.

Cognition drops a benchmark for mergeable AI code: The maker of Devin just released FrontierCode, which tests whether a human maintainer would actually merge an AI's pull request. Built with more than 20 open-source maintainers, it grades scope, style, and test quality, not just whether the code runs. Cognition says even the top performer, Claude Opus 4.8, scored only 13.4% on the hardest tasks, with rival models trailing far behind.

Apple hands Siri's AI brain to Google's Gemini: At Tim Cook's final WWDC, the iPhone maker admitted Siri needed work and confirmed Google's Gemini models will power the revamped assistant alongside next-gen Apple Foundation Models. For app developers, the bigger story is reach. iOS 27 will run on every iPhone back to the 11, the widest rollout Apple has ever attempted. Additionally, shortcuts also gain natural-language building, so anyone can wire up automations in plain English.

Your dev stack got an AI upgrade everywhere except the input layer. You're still typing every prompt, every ticket, every review comment by hand.

Wispr Flow closes that gap. Dictate into Cursor, VS Code, Slack, Linear, or anywhere else you work. It's syntax-aware: camelCase, snake_case, acronyms, and file names all come through clean. Mention a file in Cursor or Windsurf, and it auto-tags.

It's the voice layer for an AI-native workflow. Speak your intent. Your tools do the rest.

Available on Mac, Windows, iPhone, and Android. Used by millions of developers, including teams at OpenAI and Mercury.

Try free

One turn at a time. For two years, engineers have been stuck in a cycle. You write a prompt, read what comes back, and then reply. This worked for small tasks. But manual prompting hit a wall as projects got more complex. Frontier labs changed their approach early. Today, Claude writes 80 percent of Anthropic's internal code. Their latest models sustain autonomous runs for 30+ hours. The bottleneck is no longer the AI. It is the old way of working one step at a time.

Step out of the middle. When outputs fall short, most people write longer prompts. That's a trap. As OpenClaw creator Peter Steinberger and the head of Claude Code Boris Cherny point out, the highest-value move is to remove yourself from the middle. Instead of prompting the model, you write the loops, the small programs that prompt the agent, check its output against tests, and iterate until the job is done. The model becomes a subroutine.

How to start this week. You don't need a multi-agent system. What you need is to focus on automating repetitive tasks that have a clear pass/fail test.

The catch. Model makers love autonomous loops, but remember they bill by the token. A loop running while you sleep can get expensive fast. Gergely Orosz from The Pragmatic Engineer is direct about it: unless you have an unlimited budget, most developers don't really need this yet. Use loops for repetitive work if you can afford it, but always set limits. Make sure every loop has a clear gate and a strict stop condition.

When an agent run hits a dozen files across different areas and shows up as one massive diff, reviewing it feels like archaeology. Cursor’s VP of Developer Experience shared a simple fix: have the agent commit its work in smaller chunks before you ever open a pull request.

Before you push, just ask the agent to group the changes by logic and commit them separately:

The agent breaks down each change into its own commit with a specific message. This way, the reviewer can walk through the logic one step at a time instead of trying to untangle everything all at once.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

How to build an Agentic RAG project with Claude Code: You’ll learn to combine Claude Code, Agentic RAG, and the Model Context Protocol (MCP). The tutorial shows how to use Claude Code's progressive disclosure features to turn an AI agent into a digital twin, allowing for autonomous web searching and dynamic knowledge retrieval.

Browserbase: A platform for building and deploying agents that browse and interact with the web just like humans. With a single API key, your agent gets access to headless browsers, web search, page fetching, functions, model access, and identity management.

Goose (48.3K ⭐): Agent that runs locally on your machine. It features a full CLI for terminal workflows and an API for seamless integration anywhere. Built with Rust, it's designed for peak performance and portability.

How agents reshape knowledge work (by Perplexity): This study explores how autonomous AI agents are changing the nature of knowledge work. It finds that these tools shift humans into supervisory roles, enabling them to complete broader and more complex tasks faster and cheaper.

Find out how devs are now using autonomous scripts to run recursive loops directly inside the Codex app.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team