On this page
Key takeaways
- An agent correction targets one recurring mistake and reaches the agent only when its trigger matches.
- It has two parts, the guidance and the trigger that decides when the guidance applies.
- A system prompt line reaches every request, and a skill depends on the model choosing it. A correction reaches the agent when its trigger matches.
An agent correction is a short instruction aimed at one recurring mistake, delivered to your AI agent only when its trigger matches. It has two parts: the guidance, which says what to do instead, and the trigger, which decides when the guidance applies. It’s also called a scoped correction, because it’s limited to the situations you choose.
What does a correction look like?
Say an insurance agent keeps changing coverage before it checks the new driver’s license.
- Guidance: Check the driver’s license before you change coverage.
- Trigger: Conversations where the customer is adding a driver.
Conversations the trigger doesn’t match run on the original prompt, so a customer updating a billing address shouldn’t get that instruction. Preview is how you check.
How does a correction work?
These are the steps in Mirrors, where parts of the dashboard call a correction an override.
- Find the mistake. Inspect the conversations and tool results behind a recurring failure.
- Write it. You or your coding agent write the guidance and scope the trigger.
- Preview it. Review the stored conversations it would have reached, including exclusions and conflicts. Preview shows where it applies, not whether it helps.
- Activate it. The SDK returns a correction only while the agent is in active mode. Shadow and off modes return no context.
- Measure it. A holdout keeps a share of matching sessions without the correction, so you can compare outcomes using the metrics your team tracks.
How is a correction different from other fixes?
| Fix | When it reaches the agent | Who decides | Best for |
|---|---|---|---|
| System prompt line | Every request | You, once | What’s true in every conversation |
| Skill | When the model decides it’s relevant | The model | Optional procedures and knowledge |
| Guardrail | When an input, output or action is checked | Your code or policy | Limits that must hold every time |
| Agent memory | When the agent loads it | The agent | Facts carried over from earlier runs |
| Eval | Never at runtime | Your test suite | Checking a change before release |
| Correction | When its trigger matches | The trigger you wrote | Addressing one recurring mistake in one situation |
When shouldn’t you use a correction?
- The rule must hold every time. Enforce it in code, since a correction is still an instruction the model reads.
- A tool is confusing or broken. Fix the tool’s description, schema or result.
- It applies to every conversation. Put it in the system prompt.