GitHub Copilot code review now reads AGENTS.md
GitHub Copilot code review can now use a repository-level AGENTS.md file, giving teams a direct way to shape automated pull request feedback with project-specific expectations.
GitHub Copilot code review now supports repository-level AGENTS.md files. The change means Copilot can read project instructions from the repository root and use relevant guidance when generating pull request review feedback. GitHub says the update is generally available, alongside UI changes that make Copilot reviews easier to request on draft pull requests and quieter in PR timelines.
Key takeaways
- Copilot code review can now read
AGENTS.mdfrom the root of a repository. - Existing
AGENTS.mdfiles are picked up automatically as part of the code review workflow. - Teams can use the file to encode build, test, style, architecture, and review expectations that should shape automated feedback.
- Draft pull requests now show a Copilot request button in the reviewer picker.
- GitHub also collapsed some Copilot review events in pull request timelines to reduce review noise.
Practical LinkLoot angle
For teams already using coding agents, this makes AGENTS.md more than local assistant setup. It becomes a review contract: the same repository guidance that helps an agent edit code can now help Copilot judge whether a pull request fits the project.
| Surface | Best use | Limitation | Source |
|---|---|---|---|
AGENTS.md | Repository-level agent instructions for code review context | Keep it concise enough that the important rules are not buried | GitHub Changelog |
.github/copilot-instructions.md | General Copilot repository instructions | Applies beyond code review, so avoid review-only wording when possible | GitHub Docs |
| Draft PR Copilot review | Early feedback before a pull request is marked ready | Still needs human judgment on architecture, risk, and product intent | GitHub Changelog |
The useful workflow is simple: put the rules reviewers repeat every week into AGENTS.md, then request Copilot on draft PRs before asking humans for final review. Good candidates include required test commands, migration rules, security-sensitive directories, lint conventions, generated-file policies, and examples of feedback the team does not want. Keep secrets, private URLs, and one-off planning notes out of the file because repository instructions are meant to be operational context, not a private scratchpad.
Teams that already maintain /guides/ai-agent-tools style playbooks should treat this as a reason to tighten repository instructions, not to expand them endlessly. The best agent files are specific enough to change behavior and short enough to audit.
What to verify before you act
Check whether your repository already has an AGENTS.md file at the root. If it does, review it before enabling heavier Copilot review usage because old local-agent guidance may now affect PR feedback.
Confirm which instruction files your team uses today. GitHub's docs also describe repository custom instructions and task guidance such as .github/copilot-instructions.md, path-specific instruction files, and agent files. If several files overlap, decide which one owns code review rules so reviewers do not debug contradictory guidance.
Also check billing and policy settings before scaling this across an organization. Copilot code review may carry plan-specific usage costs, and organizations may have controls for agentic features, model access, content exclusion, and review activity.
Yes. GitHub's June 18, 2026 changelog says Copilot code review now reads repository-level AGENTS.md files from the repository root.
For more agent workflow patterns, see LinkLoot's guide to AI agent tools.