Code review has become the bottleneck because code generation is moving faster than teams can verify its effects.
2
A one-line data model change can pass builds, tests, and linters while breaking web, mobile, CLI, and MCP clients in other repositories.
3
Review agents need software-graph context, custom rules, ticket requirements, and independence from the agent that wrote the code.
Summary
Cole Gannaway argues that software teams now need to improve the review and verification layer around increasingly fast code generation. He uses a one-line Bitwarden data model change to show how a change can pass builds, unit tests, and linters while affecting connected web, mobile, CLI, and MCP clients. Qodo's approach builds a software graph across repositories and maps the actual integration points, so reviewers can see the likely blast radius. Gannaway also describes turning senior-reviewer judgment and past PR comments into custom rules, then supplying that knowledge before code is generated. He recommends a separate review agent because the implementation agent has different tasks, limited attention, and possible model or harness bias. Reviews can also compare changes with the original ticket and acceptance criteria, then guide an agent during implementation. The talk is practical, though its examples focus on the workflow Qodo provides.
Review has become the bottleneck as code generation speeds up
Gannaway says customers want to move from a ticket to production, but they are struggling at the pull request and review stage. Code generation is moving faster than teams have previously seen, while review remains the trust and verification layer. He says humans, coding agents, and review agents need as much context as possible to make sound decisions. The other parts of the software factory still matter, but his talk focuses on review because that is where teams are having difficulty keeping up.
A passing one-line change can still affect several clients
The example is a one-line rename in the Bitwarden data model. A developer could reasonably approve it after seeing a successful build, passing unit tests, and passing linters. Gannaway says this would miss the change's blast radius. The server is connected to web clients, mobile or iOS devices, MCP servers, and the CLI. A change in one repository or component can therefore create a production bug elsewhere, even when the usual checks are green.
A software graph maps integration points across repositories
Qodo builds a software graph of the repositories a team works on and creates an integration surface between them. Gannaway says it is not enough to know that a Bitwarden server connects to Bitwarden Android. Reviewers also need to know which actual components are connected. This graph gives coding and review agents information about dependencies and downstream effects. It lets the review process flag a large blast radius before a bug reaches production.
Custom rules turn reviewer judgment into reusable knowledge
Some review decisions are not purely functional. They depend on senior judgment, taste, or lessons from earlier mistakes. Gannaway gives a rule for public API changes: add a deprecation flag or subject the change to more review. The goal is to bring the human reviewer back into the code intelligence layer instead of making that person repeat the same check on every pull request. These rules can encode organizational knowledge for automated use.
Review context should reach the agent before code generation
Gannaway wants downstream effects, custom rules, and repository knowledge available before code is written. He mentions using an MCP server and code-base knowledge to give a coding agent this information early. That does not prevent every mistake, but it gives the agent a better chance of making the right change on its first attempt. Earlier context can also allow teams to catch a problem during code generation rather than waiting for the pull request.
The implementation agent should not grade its own work
Gannaway separates implementation from review because the implementation agent is exploring the code base, debugging, and generating code at the same time. A review agent has a more focused context, with room for information specific to the current review. He also wants the two processes to be independent. Teams could vary the underlying model, use different harnesses, or add adversarial reviewers. Agreement between independent reviewers is a stronger signal than one agent judging its own output.
Postmortems often produce lessons about process or organizational changes, but Gannaway focuses on bringing code-level lessons back into automated review. In his example, Qodo suggests a rule based on a comment on a pull request. That rule can then apply to future pull requests. The software graph adds another signal by showing that the earlier one-line change had a large blast radius, prompting a human or another agent to inspect it more closely.
Ticket requirements can guide an agent during implementation
In the Q&A, Gannaway shows how a review can link a pull request to a Jira or Linear ticket, inspect its acceptance criteria, and flag requirements gaps. He says teams can run this review during implementation, around the 25%, 50%, and 75% points. A specialized review agent can act like a manager agent, warning the implementation agent when it is going in the wrong direction or over-engineering the solution. The review checks whether the code fulfills the original ticket, not only whether it compiles.