Mirrors

Mirrors is a testing infrastructure for AI agents. It creates staging environments by rebuilding the external systems your agents interact with, then replays real production sessions against prompt, tool, or model changes before you deploy. The idea is to catch regressions like duplicate refunds, incorrect outputs, or broken tool calls in a safe simulation rather than discovering them in production.

The problem it solves is one that teams building agents know well. Your agent talks to APIs, databases, payment processors, and other systems that don't have test modes. You can't just spin up a sandbox version of Stripe or your internal inventory service to see what happens when you tweak a prompt. Mirrors rebuilds those systems as testable copies based on recorded production traffic, so you can replay interactions without touching the real thing.

The workflow starts with instrumentation. You add their SDK to your agent code, and it collects production sessions. When you're ready to test a change, Mirrors replays those sessions against your new code and shows you per-call diffs between what production did and what your updated agent would do. If a refund that should have happened once now happens twice, you see it before it ships.

It's built for engineering teams working with LangChain, LangGraph, OpenAI, or Anthropic SDKs who need to test agents against systems that lack official test environments. If your agent does anything destructive like processing payments, updating records, or sending messages, this is the category of tool that lets you validate changes without risking real consequences.

What sets it apart is the replay-and-diff approach. Plenty of testing tools mock external calls, but Mirrors records real interactions and uses them as the source of truth. You're not guessing what the agent should do. You're comparing against what it actually did in production.

Pricing is usage-based with a free tier. You get 60 replay-minutes per month at no cost, then pay $0.20 per additional replay-minute. Enterprise plans offer volume discounts, SSO, role controls, and options for on-premises or bring-your-own-cloud deployment. There's also a public API, CLI, MCP server, and SDKs in Python, TypeScript, and Go for integration into CI pipelines.