GitHub Copilot Lite Switches to Multi-Agent Reviews and Lets Reviewers Run Builds and Tests

GitHub has updated the execution architecture of Copilot code review: Lite workloads are no longer handled by a single agent inspecting a pull request. Instead, multiple agents analyze it separately, and their findings are merged into a single review. GitHub says that, in internal experiments, the number of high-, medium-, and low-severity findings ultimately addressed by developers increased by an average of 47%, 31%, and 11%, respectively, while the cost per review fell by about 8%. The company has not disclosed the number of agents, model mix, sample size, or statistical intervals, so these figures are better viewed as signals from a product experiment than as a reproducible model benchmark.

The change in execution permissions has greater technical implications. Review agents, which previously focused primarily on reading files, can now use the Copilot SDK’s full suite of shell tools to run builds, tests, and specified scripts behind the Copilot agent firewall, as well as retrieve data from approved tools or APIs. Copilot’s agentic review capabilities use a GitHub Actions runner to gather full project context. If the Actions workflow fails, a review is still generated, but it falls back to a mode with fewer capabilities. Organizations that disable GitHub-hosted runners must also configure a self-hosted runner to retain these capabilities.

The user interface now supports a more closed-loop remediation workflow. When a developer applies a Copilot suggestion, the system generates a commit message based on the changes. If a subsequent commit is determined to have addressed the original finding, Copilot automatically closes that discussion thread during re-review; only unresolved findings remain open. This can reduce manual cleanup, but it also means the same product both raises an issue and determines that it has been resolved. Regulated projects should retain audit logs for auto-resolve actions rather than treating a closed status as equivalent to human verification.

Engineering teams must also reassess the execution surface for untrusted pull requests: tests, build scripts, and tool output can all carry malicious behavior or prompt-injection payloads. Lite reviews typically consume the equivalent of about $0.05 to $1 in AI credits, with GitHub Actions minutes billed separately. The next questions to watch are whether GitHub discloses details about its ensemble method, false-positive rates, and isolation mechanisms, and whether enterprises gain more granular policies for forked pull requests, secrets, and network access.