Nine Microsoft engineers merged 1,211 pull requests in one day

Summary

• Microsoft says nine engineers merged 1,211 pull requests into main on October 21
• Agents wrote most of the code; humans owned specs, architecture, security limits and production
• The documented numbers offer a copyable operating model for small agent-assisted engineering teams

The latest

Nine engineers merged 1,211 pull requests into the main branch in a single day on October 21, Microsoft said in a November 2, 2026 post on its Command Line blog. The write-up, by partner engineer Bas Brekelmans, documents the build of Copilot Cowork from an empty repository on January 13 to global general availability on June 16.

Details

• The throughput curve: Merged pull requests climbed from 257 in January to 2,754 in May, according to the post. The busiest pre-GA week carried 783. During the week of October 12, the team merged 4,581 — roughly 650 a day — with a rising share generated by an internal factory tool.
• The division of labor: Brekelmans wrote that agents produced most of the code and proposed changes, while humans wrote or approved specifications, set architecture and security boundaries and judged results. "Agents can propose changes, but they can't decide what's allowed to happen," he wrote.
• Human checkpoints: Deploying to production and changing configuration each require a person, per the post. Critical changes always get at least one human review, with reviewers concentrating on the highest-risk changes, production impact and cases where automated checks or reviewers disagree.
• The review layer: Several specialized automated reviewers examine every pull request: one for call graphs and tests, one for operational issues, one reading every modified file, and one comparing behavior against the replaced code. Findings are sorted into blocking — correctness, security, data loss, untested mechanisms — and advisory. Critic agents run on a different model family than the agent under review.
• Python to Rust: The migration ran from September 23 to October 10. A first internal attempt was reverted on September 23; Rust served its first production turns a day later; the Python runtime was deleted on October 9; the last server cluster moved on October 10. Deletion removed roughly 8,000 files and 4.9 million lines.
• The latency case: Over a 12-hour window in late September, median time to first token on the first turn was 7.9 seconds on Python and 14.5 at the 90th percentile, versus 2.7 and 6.2 on Rust, according to Microsoft. Brekelmans noted that moving the agent loop outside the container contributed.
• Legacy failures: Microsoft recorded event-loop stalls of up to 11.9 seconds, 16.7% of session setups taking exactly 5.000 seconds, and 539 out-of-memory terminations in one week on a single server cluster.
• Dependency cleanup: A Rust workspace inventory on August 20 found 86 crates, a 13-level dependency chain, and a base-crate change forcing 43 of 63 crates to rebuild. By October 7 the workspace held about 176 crates with maximum depth cut from 10 to 3, and cross-dependency exceptions fell from 216 to zero.
• Guardrails on tooling: A test-speed tool began opening pull requests in mid-September and has had roughly 300 merged; merging requires a measured speedup of at least 5%. Modified test files dropped from about 17,700 seconds of CI time to roughly 1,700. The merge agent stops after 20 rounds or three identical CI failures.

Background

Copilot Cowork reached its first enterprise customers through Microsoft's Frontier program on March 30, about two and a half months after the first commit, and went generally available worldwide on June 16. Roughly 30 rules files govern the setup, the oldest dating to August 6.

Between the lines

The headline figure measures merge volume, not shipped product value: the same post records 10 documentation-quality passes run and only 3 kept, and a factory example where 2,148 payloads were rejected across 688 conversations. Brekelmans argued the agent was rarely the bottleneck, which locates the constraint in review capacity and dependency structure — the two areas the team spent months rebuilding.

What's next

Watch whether the factory tool's share of weekly merges keeps rising past the 4,581 recorded in the week of October 12, and whether Microsoft publishes post-migration production figures now that the Python runtime is fully removed.

Source: Microsoft Command Line blog