37signals stops writing code by hand

FROM OUR PARTNER

Aligned shows you the other half. See what buyers actually do between meeting. See what they open, revisit, and share; so you know the exact next step to move things forward. And your champion gets everything they need to win internally when you're not in the room.

Win More Deals With Aligned

TODAY AI BRIEF

3 stories worth your next 90 seconds

1. DHH declares the end of hand-written code at 37signals. In his Rails World keynote in Austin, David Heinemeier Hansson said AI agents are now the primary producers of code at 37signals, with engineers defining outcomes and reviewing results instead of typing them out. He said he generated 150,000 lines in August against a historical average of roughly 30,000 lines of Ruby a year. The twist: the HEY email backend is being rebuilt in Rust, a language he once called unpleasant for humans to write, with early back-of-the-envelope estimates of about 99 percent less CPU.

2. An OpenAI safety lead resigns and says the deployment model guarantees failures. David Robinson, who oversaw safety reports for 12 frontier model launches and drafted OpenAI's Preparedness Framework during three and a half years there, laid out his reasons in The Atlantic on October 3. He argues that shipping iteratively and fixing problems as they surface guarantees periodic failures, and cites a case where a monitoring system alerted staff but did not shut a model down as it was supposed to. OpenAI confirmed the departure and separately dismissed three staff on October 2 for mishandling sensitive information.

3. A startup deletes 800,000 lines of unit tests to make room for agents. Sherwood Callaway said his team removed 800K lines of unit tests from the Sazabi monorepo, arguing that tests lock in sloppy AI-written code, slow down agents in CI, and burn tokens on useless runs. He promised to report back with results. Robert Martin, who wrote Clean Code, replied that this is a 20-year-old argument, that integration tests leave too much code unchecked, and that test code is the vice that keeps production code stable.

Bonus read:  Anthropic engineer Thariq Shihipar measured what Claude Code's effort setting actually buys. Max effort used roughly three times the tokens of low, and lifted pass rates on security tasks from 64 to 87 percent and on hardware tasks from 34 to 75 percent, but barely moved operations tasks. His advice: draft on low, verify on high.

AI FORWORD

When agents write the code, the old rules of engineering get renegotiated

The language. Rust is fast and safe but famously hard to write, which is why most web companies never touched it. DHH's bet is that difficulty stops mattering when an agent does the typing, so you pick the language for how it runs rather than how it feels.

The tests. Unit tests exist so humans can change code without breaking it. Sherwood Callaway's experiment asks whether that still holds when the code is cheap to regenerate and the tests are what slow the agent down. Martin's counter is that the tests are the only thing holding the shape of the system steady while everything else churns.

The budget. Shihipar's numbers put a price on judgment. Three times the tokens bought big gains on tasks with hidden edge cases and almost nothing on tasks where the first approach has to be right, so effort is now a dial you tune per task rather than a fixed cost.

The oversight. Robinson's essay is the same argument one level up. If you ship fast and fix as you go, you are choosing a failure rate, and the monitoring that is supposed to catch the failures is itself just more code someone decided was good enough.

The reframe. Every one of these stories is about moving a human check from the middle of the process to the edges. That can work, but only if the edges are solid. The question for any team this week is not whether to let agents do more, it is which checks you are quietly removing when you do.

PRESENTED BY :

AI will write 90% of code by the end of 2026, and only 10% of developers will stay relevant.

The engineers in that 10% aren't smarter. They just know how to use AI.

We put together the exact playbook Anthropic engineers use to make sure you're in the top 10%.

Sign up for The Code and get access to:

Join 250K+ engineers

SOCIAL SNAPSHOT

What people are actually watching and sharing

Claude agents in Minecraft. Blendi's AgentCraft is an open-source harness where a team of Claude agents plans and codes in real git worktrees inside a Minecraft world, and an agent walks over to your character when it needs approval (990K views).

Europe's small bird. Aleph Alpha released Kolibri, a 78 billion parameter open-weight model with 3.46 billion active and a context window of up to 1 million tokens, under Apache 2.0. The weights are on Hugging Face.

Spreadsheet predictions. Prior Labs' TabPFN-3.5 takes raw tabular data, missing values and all, and returns predictions in a single pass, for jobs like churn, fraud and demand forecasting. The company reports a 99 percent win rate over classic machine learning on TabArena, which is its own figure.

Call your codebase. Armand Du Plessis showed a vibe-coded setup where he picks up a landline office phone and talks to a coding agent that edits his code (634K views).

2,500 pull requests. An Agent podcast episode with SpaceXAI engineer Lauren Tan walks through how she says she shipped 2,500 PRs in a single month with agents, and what her review process looks like.

PROMPT STATION

Turn your repeated corrections into standing rules

If you keep telling an AI tool the same thing, that is a rules file waiting to be written. This prompt works in Claude, ChatGPT or Gemini: paste in the corrections you have made over the past few weeks and it hands back a short set of standing instructions you can drop into a project, a custom GPT or a CLAUDE.md file. Try it before the next time you fix the same mistake twice.