Mirrors

Teams building AI agents can see problems in their traces: where agents get stuck, what corrections work, which context leads to good outcomes. Today those lessons stay buried in logs. The common fix is to add another instruction to the global system prompt, but that means every conversation sees the same bloated instructions regardless of relevance, and customers keep explaining the same mistakes because fixes are not targeted to the situations where they actually apply.

Mirrors is a developer tool that lets teams write guidance based on what they see in agent sessions, then routes that context into the agent mid-conversation only when specific conditions match, with no redeploy or code change needed. Before going live, users preview which sessions a new context rule would reach, including unexpected matches, and after activation they compare outcomes with and without the change using their own outcome labels and feedback. Beyond prompt guidance, teams can edit tool descriptions, tighten input schemas, or rewrite model-facing tool results. Routing decisions run locally without a network call or model inference, and if no valid configuration is available the agent runs without new context edits. It integrates with common agent frameworks (LangChain, LangGraph, LlamaIndex, CrewAI, AutoGen, and others) and coding tools like Claude Code, Cursor, and Windsurf.