GitHub published a set of Copilot code review customization and governance updates on July 17, 2026. The changes cover four areas: where custom instructions come from, how the runtime is prepared, how network access is restricted, and how code review runners are configured separately from cloud-agent runners. Together, they move the review agent toward an execution environment that can be shaped by repository and organization policy.
Copilot code review now reads custom instructions from the pull request's head branch rather than only from the base branch. Supported sources include copilot-instructions.md, instruction files, agent skills, and AGENTS.md. GitHub also says the agent will read REVIEW.md, GEMINI.md, and CLAUDE.md, allowing repositories with existing tool-specific guidance to reuse it.
Reading instructions from the branch makes it possible to test review guidance in a feature branch before merging it. It also creates a governance boundary: pull-request content should not be able to expand the agent's authority arbitrarily. For untrusted external contributions, organizations still need explicit limits on readable secrets, executable commands, and reachable network resources.
The update adds a repository-level copilot-code-review.yml file for preparing the runtime. Teams can install dependencies, configure a runner, or perform other setup needed for a review. If that file is absent, GitHub says the agent can fall back to an existing copilot-setup-steps.yml file where applicable. This separates code-review preparation from the setup used by the cloud agent.
Copilot code review now runs behind a firewall by default, restricting network access during a review. The firewall can be configured independently at repository and organization level. GitHub notes that self-hosted runners do not currently support this firewall, so reviews on those runners continue to run as before. That exception belongs in the organization's threat model rather than being hidden behind the default setting.
GitHub has also split the organization runner configuration for Copilot code review and Copilot cloud agent into separate sections on the Runner type page. This allows different runner choices, but it also means administrators must review two execution paths, their tool permissions, and their costs independently.
The practical value is not that the agent can approve more code automatically. It is that the conditions under which it reviews code are more visible, testable, and isolatable. A sensible rollout starts by making instructions, dependencies, network access, secrets, and runner choices reviewable, then uses test pull requests to verify that the agent does only what it is meant to do.
