Overview
How might we design self-correction so people can trust and act on AI output?
When to use
- Essential for autonomous AI agents, coding assistants, and automated workflow tools where self-recovery from errors improves reliability and user trust.
When to skip
- One-shot answers where retries would surprise the user with extra cost.
- Domains where automatic retries can amplify damage (notify and stop instead).
- Creative tasks where “wrong” is subjective and user taste should drive edits.
Rules
Endless retry loops with no budget or user cancel.
Corrections that hide the original failure evidence.
Claiming success after a partial fix that still fails checks.
Burning tokens on retries without showing cost impact.
Evidence
| Product | Implementation |
|---|---|
| Cursor | Agent retries after command failures with updated plans. |
| GitHub Copilot | Iterative fix suggestions when prior code fails. |
| AutoGPT-style agents | Loop on errors with revised steps in the task log. |
| LangChain agents | Tool error handling and re-planning in traces. |