Contain before correcting
Do not let an agent keep expanding its changes after the task boundary has been violated. Pause the work so review has a stable diff and evidence set.
Answer page
Stop the task, preserve the evidence, compare the result to the declared boundary, and create an explicit corrective task or revert through normal review controls.
Direct answer
When an agent makes the wrong change, do not hide the failure in a new prompt. Stop the task, inspect the diff and validation evidence, decide whether to revert or correct, record the cause, and create a newly scoped follow-up task with the right acceptance criteria.
Practical guidance
Do not let an agent keep expanding its changes after the task boundary has been violated. Pause the work so review has a stable diff and evidence set.
Revert, amend, or create a corrective pull request through the same version-control and review practices used for human-authored changes.
Identify whether unclear scope, missing validation, or ambiguous product intent caused the error. Put the correction in the next task definition rather than relying on a stronger warning prompt.
Verified demo evidence
$ day-shift reconciliation build --milestone-overview <milestone-overview>
reconciliation_status: completedAuthorship and sources
Keep exploring
Clear triggers for moving coding-agent work to accountable human judgment.
Read nextUse declared target paths and escalation to prevent hidden scope expansion.
Read nextThe objective, scope, acceptance criteria, and validation a coding agent needs before changing code.
Read nextFind your answer
Search the answer pages, concept, comparisons, demo, and FAQ without leaving the site.
Search by a problem, workflow step, or tool boundary.