The right correction. Only where it applies.
Mirrors adds it only in the situations you choose.
Backed by
CombinatorFind recurring agent mistakes.
CombinatorFind recurring agent mistakes.
Fix them in real time.
Spot the mistakes that keep coming back, correct them while your agent works, and measure whether each correction helped.
works with any agent framework
- LangChain
- LangGraph
- LlamaIndex
- CrewAI
- AutoGen
- Pydantic AI
- smolagents
- Google ADK
- Hermes
- AI SDK
- Mastra
- OpenAI
- Anthropic
- MCP
- Claude Code
- Codex
- Cursor
- Windsurf
Your system prompt is becoming a history of everything your agent got wrong.
Your agent makes a mistake. You add a line to the prompt. It makes another. You add another.
Soon every run carries dozens of patches, and no one knows which ones help.
How it works.
From recurring mistake to measured correction.
01
Find recurring mistakes.
See what your agent keeps getting wrong in production.
02
Write a correction.
Say what it should do instead.
03
Choose where it applies.
Preview the past conversations it would have reached.
04
Measure the result.
Compare outcomes with and without it.
See what went wrong. Correct it where it belongs.
Every correction in one place.
Scoped to accounts and workflows. Update one, not your whole prompt.
Fewer repeat mistakes. Happier customers.
Compare outcomes with and without each correction.
Fewer patches. Fewer repeat mistakes.
Before and after Mirrors.
TODAY WITH MIRRORS
01Every mistake adds a line to the system prompt.
↗Each correction lives in one place.
02Every patch runs on every request.
↗Corrections apply only where they belong.
03Changing requirements means editing the whole prompt.
↗You update one correction.
04No one knows which patches help.
↗Every correction is measured.
711,933 agent sessions. And counting.
Find what repeats, correct it where it happens, and measure what changed.
See what that looks like for your agentFAQ
What does Mirrors do?
Mirrors finds the mistakes your AI agent keeps making in production and lets you fix each one with a correction: a short instruction your agent gets only in the situations where it applies. You preview a correction before it goes live, then compare outcomes with and without it.
Why not add the fix to my system prompt or a skill?
A line in the system prompt reaches every conversation, so fixes pile up and your agent carries all of them, even when it’s doing something simple. A skill is read only when the model decides it’s relevant, so the fix can be missed. Mirrors adds each correction when its situation comes up, keeps every correction in one place, scoped to the right accounts and workflows, and shows whether it helped.
How is Mirrors different from evals and observability tools?
Observability shows what your agent did. Evals check it against test cases before you ship. Mirrors acts on what keeps going wrong in production: it surfaces recurring mistakes, applies your corrections where they happen, and measures each one on real runs. It works alongside both.
How do I connect my agent?
Install the Mirrors SDK, send your agent’s conversations, and request corrections once per user turn. It works with any agent framework, and your app keeps control of its model calls and tools.
Does Mirrors add latency?
The SDK waits at most 750 ms per user turn, and you can set a shorter limit. If Mirrors doesn’t answer in time, your agent continues without corrections.
Can I check a correction before it goes live?
Yes. Mirrors previews each correction against your stored conversations, so you see every one it would have reached, including any you didn’t expect, before you turn it on.
Fix recurring AI agent mistakes with real-time guidance.
Book a demo or email devs@runmirrors.com.
Book a demo