Code Review Now Powered by Agentic Architecture
GitHub CopilotGitHub Copilot's code review feature has been rebuilt on an agentic tool-calling architecture that proactively gathers repository context before surfacing feedback on a pull request. The redesign has produced an 8.1% increase in positive developer feedback and is now generally available for all paid Copilot subscribers.
Sources & Mentions
3 external resources covering this update
Copilot Code Review Gets an Agentic Overhaul
GitHub has rebuilt Copilot code review on a new agentic tool-calling architecture, marking a fundamental shift in how the AI assistant evaluates pull requests. As of March 5, 2026, this upgraded system is generally available.
From Isolated Review to Repository-Aware Reasoning
The previous architecture reviewed code in isolation. The new agentic design uses tool calling to retrieve relevant context: it can explore directory structure, look up related code, cross-reference similar patterns elsewhere in the repo, and understand architectural invariants.
Quality and Signal Improvements
- Higher-quality findings that prioritize correctness and architectural integrity
- Lower noise — Copilot stays silent in 29% of reviews where it finds nothing actionable
- More actionable guidance
- 8.1% increase in positive developer feedback
When Copilot does leave feedback, it averages 5.1 suggestions per review, clustered into logical groups.
Scale and Adoption
Copilot code review usage has grown 10x since its April 2025 launch. GitHub's AI reviewer has now processed over 60 million code reviews and handles more than one in five pull requests on the platform. Over 12,000 organizations run it automatically on every PR.
Infrastructure
Agentic code review runs on GitHub Actions. Organizations that have disabled GitHub-hosted runners must complete a one-time self-hosted runner configuration.