Microsoft built Copilot Cowork, an agentic execution layer for multi-step workflows. Nine engineers merged 1,211 pull requests in one day. Agents wrote most of the code but humans defined work and judged results. The team moved the live runtime from Python to Rust while active. Six practices emerged for specs, boundaries, checks, learning, parallel work and loop limits. Copilot Cowork reached worldwide GA on June 16, 2026. Frontier program launched March 30 with first enterprise customers. The Rust migration ran September 23 to October 10. Agents run in containers without user credentials. A trusted host handles identity, policy, approvals and tools. Reviewer agents now arrive within minutes of a PR. Human review remains for critical changes. Checks must prove they ran and include tests to fail on purpose.
What changed
Agents wrote code without specs, leading to rework. Reviews were human-only and became a bottleneck.
What it unlocks
Teams can run dozens of agents in parallel with clear boundaries and automated checks.
- 9 engineers
- 1,211 pull requests in one day
- 257 to 2,754 merged PRs per month (Jan to May 2026)
- 783 PRs in busiest week before GA
What you need to act on it
- a markdown spec written before code
- narrow tasks per agent
- trusted host for identity and policy
- human review for critical changes
Sources