The latest articles on DEV Community by Baessi (@seunghunbae3svs). https://dev.to/seunghunbae3svs https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3923607%2F94ef6d00-25b8-4de1-aa8e-7c325ebf6dc3.png https://dev.to/seunghunbae3svs en Baessi Sun, 10 May 2026 17:23:08 +0000 https://dev.to/seunghunbae3svs/6-months-solo-on-a-multi-agent-pr-reviewer-1093-vs-380-blockerspr-claude-alone-on-my-20d0 https://dev.to/seunghunbae3svs/6-months-solo-on-a-multi-agent-pr-reviewer-1093-vs-380-blockerspr-claude-alone-on-my-20d0 <p><a href="https://dev.tourl"></a>TL;DR: I built a 3-LLM code reviewer (Claude + GPT-5 + Gemini that deliberate). My synthetic-bug benchmark shows 3×<br> the depth at the same catch rate as Claude alone. But 15 synthetic PRs is not enough. I need YOUR PRs to validate or<br> kill the hypothesis.</p> <p>Background:<br> 6 months ago Claude solo review kept missing things I considered blockers but it called "minor". Tried adding more<br> models in parallel + deliberation. Result on my private corpus:</p> <ul> <li>Claude alone: 3.80 blockers/PR</li> <li>3-agent council: 10.93 blockers/PR</li> <li>Both 100% catch on synthetic bugs</li> </ul> <p>Pattern, after debugging the gap: one model skips a missing test that another catches. A "minor" by Claude becomes a<br> blocker by Gemini. Single-agent has no second perspective.</p> <p>The bigger feature is PRD-aware review. .conclave/prd.md → agents flag spec deviations as first-class blockers. Scope<br> creep, route mismatches, forgotten acceptance criteria.</p> <p>What I need:</p> <ul> <li>Run on a real PR, tell me where wrong</li> <li>Compare vs your usual reviewer (Claude / Cursor / human)</li> <li>Send false positives, I incorporate into federated failure-catalog</li> </ul> <p>How:</p> <ul> <li>Demo (3 free/day): <a href="https://conclave-ai.dev/#try" rel="noopener noreferrer">https://conclave-ai.dev/#try</a> </li> <li>GitHub App + BYO key = unlimited free</li> <li>CLI: npm i -g @conclave-ai/cli</li> </ul> <p>Source-available (FSL-1.1-Apache-2.0): <a href="https://github.com/seunghunbae-3svs/conclave-ai" rel="noopener noreferrer">https://github.com/seunghunbae-3svs/conclave-ai</a><br> Stack: TS / Node 20 / Cloudflare Workers + Containers + D1 / Mastra. 26 packages, 2691 tests.</p> <p>Limitations I know:</p> <ul> <li>Beta, things break</li> <li>Cost scaling on large diffs untested</li> <li>Spec-mismatch only useful if you maintain a PRD</li> <li>I'm one person + Claude pair-programming — bus factor 1</li> </ul> <p>If the numbers don't survive contact with real codebases, I want to know. Poke holes.</p> <p><a href="https://github.com/seunghunbae-3svs/conclave-ai" rel="noopener noreferrer">https://github.com/seunghunbae-3svs/conclave-ai</a></p> agents ai llm showdev