Microsoft Engineers Detail Copilot Cowork's 4.5-Month Path to GA and Python-to-Rust Migration
·2026.11.03 00:13
Key point
Microsoft's Copilot Cowork reached GA in 4.5 months, with a record 1,211 PRs merged in a single day, while a live Python-to-Rust migration reduced p50 time-to-first-token from 7.9s to 2.7s.
Details
- Microsoft's Bas Brekelmans details the 4.5-month journey of Copilot Cowork from first commit to General Availability on June 16, highlighting a live migration from Python to Rust to resolve scaling issues.
- The Rust migration eliminated 11.9-second event loop stalls and reduced p50 time-to-first-token from 7.9s to 2.7s, with p90 latency dropping from 14.5s to 6.2s.
- Development relied on six core practices, including writing specs before code and running narrow agents in secure containers. On October 21, nine engineers merged a record 1,211 PRs into the main branch.
- An automated "factory" system processed production signals to generate work items, merging 4,581 PRs in the week of October 12, 2026, while maintaining human checkpoints for high-risk changes.
Development Methodology
- Agents operate in isolated containers without user credentials, with a Trusted Host managing identity, policy, and audit logs.
- Automated review bots compete to find issues, providing initial feedback within approximately 4.5 minutes, though critical changes still require human review.
- Approximately 30 rule files are stored in the repository to prevent repeated agent errors across sessions.
Rust Migration Execution
- Python runtime issues included 16.7% of session setups hitting a 5.000s timeout and 539 OOM kills in one week.
- Verification used JSON fixtures to ensure Rust output matched Python behavior exactly.
- The Python runtime, comprising roughly 8,000 files and 4.9 million lines, was deleted on October 9, 2026, with the final cluster cutover on October 10, 2026.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.