Agent cannot converge on a 'simple' fix: each fix surfaces another masked failure, many attempts, no clean close
Open 💬 1 comment Opened Jun 23, 2026 by Mig-Sornrakrit
Asked to fix what looked like a simple defect, the agent spent an extremely long, multi-day, many-attempt effort and still could not fully close it. The core failure mode:
- Whack-a-mole on hidden failures. Fixing the real defect exposed a chain of pre-existing masked problems — a circular/self-referential test reference, a wrong code path, a data-type mismatch in an external API call, then a value-ordering mismatch — each only revealing the next after a fix was applied and re-verified. The user experiences this as 'every time you say it's fixed, you hand me a new hidden failure.'
- No upfront map of the blast radius. The agent did not enumerate the layered dependencies before starting, so it discovered them serially across many cycles instead of surfacing them once.
- A self-inflicted regression mid-fix. At one point the agent put an edit in the wrong function (a different module's code path) and had to catch and revert it — exactly the 'new problem from the fix' the user is exhausted by.
- Effort wildly disproportionate to the stated task size. A 'simple' function consumed days and dozens of tool cycles, with the user paying for each, and the final state still has a residual the agent could not close.
- Repeated reassurance that didn't hold. Multiple 'done / it works now' moments were followed by the verifier surfacing more failures.
Requests:
- When a fix touches shared/layered code, the agent should map and report the full dependency/blast-radius BEFORE editing, so masked failures surface at once, not one-per-cycle.
- The agent should never silently keep grinding; after N failed convergence cycles it should stop, report the honest residual, and ask, rather than continuing to bill the user.
- Effort should be bounded relative to the stated task size; a 'simple' fix ballooning into days is itself a signal to stop and escalate.
- 'Fixed' should never be claimed before the authoritative check passes end-to-end.
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗