- Theo argued that when agents write code, teams need a safer merge step and small, reversible damage, with undo weighted above extra review layers.
- He conceded that Anthropic's system is not for general use, and his argument assumes a mistake can be undone.
Theo, host of the show Theo - t3.gg, argued that safe speed with AI-written code needs two things. Make the merge step safer, and keep the damage from a bad merge small and reversible. A merge adds a change to the shared codebase. He weighted undo above extra review layers, in the 16 October episode "Anthropic Rebuilt Git (yes really)".
What was said
Theo discussed an Anthropic engineering post. It says Anthropic merged about 2,500 code changes a day in May 2026 and over 30,000 a day by September. Theo said tools built for human pace strain at that volume.
His answer has two sides. Make the merge safer, so a change is likely fine by the time a person looks. Then shrink the damage of a bad merge: release to a limited set of users first, and revert if it breaks.
He gave a personal example. One night he merged more than 50 pull requests in about four hours, between nightly builds that ship every six hours. Before bed he could test the build and revoke it if it was bad. He said: "it's more important to be able to undo a bad change than to add thousands of layers in front of the changes." He added: "no matter how strict your processes are, you will ship something broken."
Why it matters
Our reading: the question for a manager or buyer shifts from how much review a change gets to how fast the team can find and undo a bad one.
If so, the cost of a miss lands on whoever does the cleanup, and on the first small group of users who meet the bug. Buyers can ask vendors who gets a release first, and how quickly it can be pulled back.
The other side
Theo conceded limits. He said the post calls Scion, Anthropic's in-house source control system, no replacement for Git in general use. Its ideas fit only huge codebases with tens of thousands of engineers or swarms of agents. He also noted that Anthropic can spend far more on AI tokens than most firms.
He said teams not moving this fast are either where a mistake could kill people, or have not built the safety tooling. The argument assumes a mistake can be undone. Where harm cannot be reversed, fast rollback does not help.
The post also asks what CI, the automated testing step, even means when many changes merge while tests run. Anthropic invested in automated testing after merge and before deploy instead. The excerpts do not say how early users in a staged rollout are protected.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- Theo - t3.gg: Anthropic Rebuilt Git (yes really) (2026-10-16)
