Microsoft VS Code PM Harald Kirschner Details Shift from Monthly to Weekly Releases via AI Agents
At the 2026 AI Engineer World's Fair, Harald Kirschner, Principal Product Manager for VS Code at Microsoft, presented how the team transitioned from monthly to weekly releases by rebuilding their system around AI agents. He outlined three stages: shipping faster (agent-ready codebases, skills), holding quality (mandatory AI review, auto-fix PRs), and learning faster (VSC-Bench evals). The talk offers practical insights for engineering teams integrating AI agents into production workflows.
ImportanceLocalEvidenceE2 unreplicatedWrite-upQuick
Harald Kirschner, Principal Product Manager for VS Code at Microsoft, shared insights at the 2026 AI Engineer World's Fair on how his team shifted from a decade-long monthly release cycle to weekly releases. This change was driven not just by increased AI usage, but by rebuilding the entire engineering system around AI agents.
Kirschner detailed three core stages: 'Shipping Faster,' which involved making codebases agent-ready, encoding expertise as skills, leveraging TypeScript Go for 10x faster builds, and allowing agents to drive the app; 'Holding Quality,' achieved through mandatory AI code reviews, AI issue triage, converting error stacks into auto-fix PRs, and staged rollouts; and 'Learning Faster,' utilizing VSC-Bench evaluations, daily prototyping, and smaller squads.
Specific mechanisms mentioned include the rise in committed agent-written code from 55% to 86%, the compounding effect of slow CI when multiple agents interact simultaneously, establishing self-correcting loops by letting agents use the application, and transforming 51 billion daily telemetry events into auto-filed issues and fix PRs. He also cited a case where a 5-character 'hello world' evaluation cost one model 70x more tokens.
This resource is valuable for readers interested in AI agent integration in production environments, engineering workflow optimization, and large-scale software release management. As there is no independently verified transcript or third-party reproduction of the claimed metrics, these points should be viewed as first-party experience sharing rather than audited results.