# Mirrors

> Mirrors (mirrors.dev) is a developer tool for AI agents. Teams turn what they see in their agent's traces into guidance, and Mirrors routes that context into the agent mid-conversation when the situation matches. Teams preview where a change would apply against stored traces, compare outcomes against holdout groups, and ship changes without a redeploy or code change. Mirrors is backed by Y Combinator.

Mirrors describes its workflow in four steps: analyze your traces, write the fix as context, route it in real time, and measure the outcome. Your team defines what success looks like using its own outcome labels, checks, and customer feedback. Mirrors does not define ground-truth success or generate recommendations.

## Product

- [Homepage](https://mirrors.dev/): What Mirrors does, the problem it addresses (lessons from agent sessions stay buried in traces, and fixes pile up in a global prompt), and how it works.
- [How it works](https://mirrors.dev/#how-it-works): Analyze your traces, write the fix as context, route it in real time, measure the outcome.
- [Dynamic context](https://mirrors.dev/#platform): Session review, previewing context against stored sessions (matches, misses and unexpected matches), mid-conversation routing, and outcome comparison. Examples on the page are illustrative.
- [Pricing](https://mirrors.dev/#pricing): No public price list. Teams start with one workflow, and pricing is discussed per use case.
- [FAQ](https://mirrors.dev/#faq): What Mirrors is, how it differs from adding instructions to a system prompt, what can be changed, how changes are previewed and measured, and whether routing adds latency.

## Key facts

- What can be changed: append guidance, edit a tool description, tighten a tool's input schema, or rewrite the tool-result content shown to the model. Tool descriptions and schemas are set at session start; guidance and result edits can apply during the conversation.
- Preview: routing conditions can be previewed against stored traces to review matches, misses, and unexpected matches before a change goes live.
- Measurement: sessions that received the context are compared with holdout groups using outcomes the team tracks. Teams keep what helps, revise what does not, or roll a change back.
- Latency: routing decisions run locally, without a network request or model inference to decide what applies. Configuration refresh and telemetry happen separately. If no valid configuration is available, the agent continues without new context edits.
- Scale: Mirrors' design partner runs over a million agent sessions through Mirrors.

## Company and contact

- [Book a demo](https://mirrors.dev/request-access): Book a call with a Mirrors founder. Bring a workflow, your traces, and a change you want to try.
- [Sign in](https://platform.runmirrors.com/login): The Mirrors platform.
- Email: devs@runmirrors.com
- [Y Combinator profile](https://www.ycombinator.com/companies/mirrors)
- [LinkedIn](https://www.linkedin.com/company/runmirrors)
- [X](https://x.com/runmirrors)

## Optional

- [Privacy](https://mirrors.dev/privacy)
- [Terms](https://mirrors.dev/terms)