You're reading one of the top pains on Cursor (AI IDE) — a free sample of the catalog. Create a free account to browse all Shopify & Atlassian pains and the basic Chrome catalog.
Cursor developers can't trust agent 'fixed' claims without proof of a failing-then-passing
.22
Demand — how loud & how many (score)
Willing to payImplicit
Momentum (7d)▼ 1%
Cluster50 posts · 38 people
MarketplaceCursor (AI IDE)
Updated5d ago
The story
The pain
We keep seeing agents say 'fixed' when they only changed the code but never reran the failing command. The diff looks right, but the bug is still there because the environment wasn't set up, the race only shows under load, or the agent just narrated the edit without verifying. We need the same session to show the exact command that previously failed and its new green output before we can trust the claim.
Evidence
What people said · 50 complaints, 38 people
I started requiring a failing command then a passing one in the same session before I trust fixed. Agents love narrating the edit. They are quieter about rerunning what broke.
The sticky version for us was: the same reply that says fixed has to paste the exact command that previously failed and its new output. No paste → write "not verified."
yeah forcing the failing command paste before any "fixed" claim is the only rule that stuck for me. half my agent sessions used to end with a clean summary and zero runs because the env vars weren't even set.
this is why i never trust the agent's own verdict. my setup: agent works in a sandbox, then i grade the result against hidden tests it never saw. "fixed" is a claim, the test suite is the verdict.
I've had an agent run the exact failing command, paste a green result, and still be wrong: the race only shows under load, the timestamp only breaks in UTC+7, the bad row only exists in prod. The run proved the code, not the bug.
Export the pain — problem, evidence and market signals — as Markdown, PDF or JSON. The Markdown export ships with a built-in briefing you can hand straight to a developer or an AI agent.