<p>Development teams are seeing a rapid increase in code contributions, many of them generated with artificial intelligence (AI) tools, including pair-programming copilots. As the volume of AI-generated code increases, human reviewers can quickly become overwhelmed.</p> <p>The instinct to throw “an AI reviewer” at the problem often does not work. A single agent produces shallow comments, misses context, and flags either too many or too few issues, depending on how it is prompted.</p> <p>In this talk, Vikram Vaswani presents a more effective pattern, built and demonstrated from scratch: three specialised review agents working in parallel, followed by a developer agent that applies the fixes. One reviewer focuses on architecture, one on security, and one on test coverage. When all three finish, a fourth agent rewrites the code to address each issue and opens a new pull request (PR) for human sign-off.</p> <p>Attendees will leave with a concrete, working pattern they can adapt to their own projects and a clearer understanding of where AI can genuinely reduce review bottlenecks.</p> <p>Benefits</p> <ul> <li>Explores multi-agent orchestration patterns rather than typical single-agent Model Context Protocol (MCP) interactions</li> <li>Provides developers with concrete patterns, including a shared file system and execution isolation, for designing swarm architectures</li> <li>Highlights failure modes in multi-agent systems</li> </ul> <p>Outline</p> <ul> <li>The review bottleneck: why the growing volume of AI-generated code is overwhelming human reviewers</li> <li>Why a single AI reviewer does not work well</li> <li>The multi-agent pattern: three specialists and one developer agent</li> </ul> <p>The architecture</p> <ul> <li>How agents can collaborate without passing large blocks of context</li> <li>Limits, failure modes, and where humans must remain in the loop</li> <li>Live demonstration</li> </ul> <p>Context<br /> This situation is real. For example, tldraw and Ghostty have stopped accepting contributions due to the high volume of AI slop. For additional context, the session references “Stop Closing the Door. Fix the House,” a blog post by Angie Jones, a maintainer of Block’s goose project, which has more than 300 external contributors:<br /> <a href="https://lnkd.in/exquiTP6?trk=public_post_reshare-text">https://lnkd.in/exquiTP6?trk=public_post_reshare-text</a></p> <p>The session also references Kelsey Hightower’s comment on the topic:<br /> <a href="https://www.linkedin.com/posts/kelsey-hightower-849b342b1_stop-closing-the-door-fix-the-house-angie-activity-7425980144877555712-Uedv">https://www.linkedin.com/posts/kelsey-hightower-849b342b1_stop-closing-the-door-fix-the-house-angie-activity-7425980144877555712-Uedv</a></p>