Mirrors documentation
Mirrors integrates with agent harnesses to record conversations and apply corrections for repeated AI errors.
mirrors.devCrawled Nov 14, 2026
Mirrors integrates with agent harnesses to record conversations and apply corrections for repeated AI errors.
Mirrors integrates with TypeScript agents at three key points.
Use shadow, active, and off modes to control optional context delivery.
Add capture, context retrieval, and completion hooks to your existing Python agent.
Scope guidance to a recurring failure and review its matches before activation.
Mirrors prepares prompts for a single interaction with an existing agent.
Notes on keeping agent corrections useful, inspecting the decisions that follow, and reducing repeated mistakes in production workflows.
Plain definitions of the terms behind Mirrors, starting with agent corrections and how they differ from prompt lines, skills and guardrails.
How corrections compare with observability, evals and the other ways teams handle recurring AI agent mistakes.
What self-improving agents do, where automatic prompt rewriting falls short, and why a review step between a lesson and live behavior is worth keeping.
How to tell when a system prompt has grown too far, what to move out of it, and how to add an instruction only in the situations where it applies.
AI agents repeat mistakes when a fix lands in the wrong place, nobody groups the failures, or nobody measures the change. How to find the pattern and make a fix stick.
Unlimited agents on every plan. Hobby is free with 1,000 turns a month. Growth is $299 a month, plus $0.005 per turn. No agent or correction fees.
Mirrors identifies recurring AI agent errors, applies corrections only where needed, and tracks their effectiveness.
How Mirrors processes website visits, contact requests, account information, agent data, and billing information.
Information about the Mirrors website, service, plans, billing, and separate product agreements.
AI agent observability tracks real-time agent actions and outputs. Evaluation tests compare agent performance against predefined benchmarks before deployment. Corrections dynamically adjust inputs for specific scenarios while the agent operates. Teams managing production AI agents typically integrate all three approaches to address distinct operational needs.
Researchers identify three key components for deploying self-improving AI agents in production: automating failure detection, generating repair proposals, and testing fixes. The system retains a human or trusted review process to validate changes before they affect live operations. Agents adapt behavior based on feedback from prior executions.
Mirrors.dev recommends avoiding overly long system prompts in AI models, as each rule competes across all requests. The solution involves retaining only universally applicable instructions in the prompt and relocating edge-case rules to more targeted systems. This approach prevents rule conflicts in unrelated conversations.
Mirrors.dev defines an **agent correction** as a targeted fix for AI agents, addressing one specific recurring error. The system uses a **trigger** to identify when the mistake occurs and **guidance** to specify the correct action. This approach limits corrections to predefined scenarios, avoiding broad overrides.
AI agents repeat mistakes because they lack built-in learning from production runs. The system relies on manual intervention to identify error patterns, adjust inputs or actions, and validate fixes. Without these updates, the same flawed behavior persists across identical scenarios.