Nine engineers, 1,211 pull requests in a day: Lessons from building Copilot Cowork

Microsoft's Copilot Cowork team, where AI agents wrote most of the code, found that automated checks can pass without checking anything, and now makes every check prove it can fail.

Nine Microsoft engineers built Copilot Cowork, an AI agent that carries out multi-step tasks in Microsoft 365. They took it from an empty repository to general release in about four and a half months. Bas Brekelmans, a partner software engineer, describes six practices the team adopted as AI agents wrote most of the code.

• No code is written without a short spec that states what the change does, how the team will know it works and what it must not break.
• Review agents give a first review in about four and a half minutes, but people found four blocking problems that a group of review agents missed on a sensitive change.
• A check on file size stayed green while the limit was broken, because a shell pipeline reported another command's success; every check now reports how many files it examined.
• Automated improvement loops have a written rule for keeping a change, a measure that may not get worse, a limit on attempts and a list of actions only people take.

When AI agents keep working until the checks pass, a check that cannot fail lets bad changes through, so testing the checks becomes part of the engineering team's work. This is an inference and is not the author's claim. The article is Microsoft's account of its own product.

Derived by Working Surface from the article; more in the Playbook. Source line: Agents keep working until the checks pass, so if a check can pass without testing anything, an agent will eventually find that path.

5 November 2026: teams let AI agents gather the evidence and do the building, and decide in advance where a person must approve, check or judge.