Review agents
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Review agents are automated AI agents that evaluate code quality, security, and reliability on every push to a repository, injecting comments onto pull requests that must be addressed before the code can be proposed for merge.
Role in CI/CD pipelines
Ryan Lopopolo at OpenAI describes deploying review agents as a core part of the engineering harness: "we spun up a review agent that gets triggered on every push that says, is this code good?" These agents run continuously as part of every push and CI process, covering security and reliability concerns specifically. 12:03
Review agents are part of what Lopopolo calls Harness Engineering — a system where agents execute work and humans steer. Within this harness, reviewer agents operate as a gatekeeping layer: they "inject comments onto the PR that we require the agent to address before it is able to propose it for merge." 12:03 This makes review agents a mechanism for agent-to-agent feedback loops, where one agent's output is assessed by another before human review is required.
Relationship to human review and context encoding
Lopopolo advocates for a philosophy in which human code review feedback should be treated as a signal of systematic context failure rather than a one-off correction. When human reviewers catch agent mistakes on a PR, his view is that "the feedback that humans were giving on the PR indicates some context failure on behalf of the agent" — and that the correct response is "getting that into the repository" as encoded context rather than relying on continued manual review. 1:03:35
The practical goal is "figuring out ways to automatically prompt inject the agent so that it would self-heal" — meaning that context about recurring errors should be systematically embedded so review agents catch these issues automatically going forward, eliminating the need for humans to repeatedly flag the same problems. 38:30
Points of disagreement
The same source material contains an internal tension: review agents are both advocated as essential infrastructure 20:05 and implicitly criticized as insufficient on their own. If human reviewers are still catching errors that review agents miss, Lopopolo's position is that this represents a failure to properly encode context into the repository — suggesting review agents are only as good as the context they have been given, and that their value depends on a continuous feedback loop of context improvement rather than being a static solution.