GitHub Copilot's Lite Code Reviews Get an Agent Ensemble
GitHub CopilotGitHub Copilot's fast, low-cost Lite code review tier now produces its reviews using an ensemble of agents instead of a single model working alone, with each agent independently analyzing the pull request before Copilot merges their findings into one review. GitHub reports the change increased the average number of addressed comments per review by 47% for high-severity findings, 31% for medium, and 11% for low, while cutting review cost by about 8%. The update pairs with a separate change letting all review tiers run the full set of shell tools from the Copilot SDK, including build and test commands, to validate findings before surfacing them.
Key Takeaways
- Lite reviews now use an ensemble of agents rather than one model working alone, with their independent findings merged into a single review.
- The switch delivered a 47% increase in addressed high-severity comments, plus gains of 31% (medium) and 11% (low severity), according to GitHub's own testing.
- Lite reviews got cheaper, not more expensive, with GitHub citing roughly an 8% cost reduction alongside the quality gains.
- Copilot's review agent, across all effort levels, can now run build commands, tests, and scripts through the Copilot SDK's shell tools to validate its findings.
- GitHub reports the shell-tool access alone led to more high-severity findings and fewer nits in developer feedback.
- Both updates are fully automatic and require no settings change; they change review quality, not the review request/response flow.
Sources & Mentions
1 external resource covering this update
Multiple Agents, One Review
Effort levels let teams choose how much reasoning power Copilot spends reviewing a given pull request, with Lite built for speed and low cost on straightforward changes. Until now, Lite reviews came from a single model working through the diff on its own. GitHub has now rebuilt Lite's review pipeline around an ensemble: several agents each analyze the same code independently, contribute their own perspective on what matters, and Copilot combines their findings into a single, unified review rather than presenting multiple competing sets of comments.
According to GitHub's own experimentation, the shift meaningfully changed review outcomes. The number of high-severity findings developers actually addressed after a Lite review rose by 47%, medium-severity findings by 31%, and low-severity findings by 11%, all while the average cost of running a Lite review fell by roughly 8%. In effect, GitHub is arguing that spreading a review across several cheaper models, then reconciling their output, delivers more reliable results per credit spent than a single model running alone, at least at the Lite tier.
Reviews Get Better Tools, Too
The same release quietly expanded what any Copilot code review, at any effort level, can do while forming its opinion. Building on the file-reading tools it already used, Copilot's review agent now has access to the full set of shell tools available through the Copilot SDK, run behind the Copilot agent firewall. That means a review can execute build commands, run test suites, run targeted scripts, or query other available tools and APIs to verify a claim about the code, rather than relying purely on static reading. GitHub says this alone led developers to leave more positive feedback on Copilot's comments, with the agent surfacing more genuinely high-severity findings and fewer low-value nits.
Neither change alters how a developer requests or receives a review; both operate entirely behind the scenes, aimed purely at making the review itself smarter and more trustworthy for the same interaction developers already know.