Claude code review is a slot machine: Addy Osmani on defining 'done' for agents
addyosmani · x · 2026-10-12
Developer Lucas complains that asking Claude to review code is like a slot machine: every pass finds brand-new bugs, and it never runs out—so how do you decide the code is done?
Addy Osmani's answer: verification is key to getting great results from your agent.
- Clearly define "done" before code review starts; without it, any reviewer (human or agent) will keep finding improvements
- Give Claude checks it can run itself: bash commands, tests, clicking through the app in a browser, and require it to pass them before claiming completion
You're finished when nothing blocks merge and ship.
More from coding & agent
- levelsio cancels nearly all SaaS subscriptions: money now goes only to AI inference, content, data centers, power and taxes — karlwaldman · 2026-10-12
- Grok Bot is winning the agentic personal assistant race, says dev — Arindam_1729 · 2026-10-12
- Does Codex refuse or cut back work when its own budget estimate runs high? — r618NecessaryStation · 2026-10-12
- Excalidraw for slides is the GOAT: MCP integration plus animations, no AI slop — HamelHusain · 2026-10-12
- Autoresearch for LLM pretraining needs humans in the loop; stacked changes are the killer — menhguin · 2026-10-12
- Non-coder shares cost-split agent pipeline: frontier models only where judgment matters — Soggy_Yogurtcloset35 · 2026-10-12