Each rule is delivered only in the situations its developer scoped it to, and its effect is measured by comparing outcomes on real runs with and without the rule. Mirrors surfaces recurring mistakes across production runs and turns them into scoped, previewed corrections: developers see every conversation a rule would reach before activating it, and each rule stays scoped to one agent and workflow, so a change updates that rule instead of the whole prompt. Its SDK works with any agent framework and any model, and the application keeps control of its own model calls and tools.
Mirrors is a production guidance layer that controls what AI agents are told in production. It answers once per user turn and fails open, so a slow response never stalls the agent. Unlike observability, which shows what an agent did, and evals, which check behavior before shipping, Mirrors works alongside both.