Mirrors

Not yet independently verified. Mirrors names no customers and publishes no pricing or security certifications; the 711,933 agent sessions figure on its site and the 75 errors example are the company's own. We have seen no independent review of how well corrections reduce errors. Neither its site nor its launch page states where it is based. We will update this when it can be confirmed, and remove this note.

When a production AI agent keeps making the same mistake, the usual remedies are another line in the system prompt or another skill file, and neither tells you whether it worked. One design partner's agent made 75 errors while editing a single travel guide, with the instruction to stop already in its prompt four times, according to the founders. Mirrors takes a narrower route.

Corrections that appear only when they apply

The product surfaces recurring failures from sessions, outcomes and user feedback. A developer writes a short correction, says which accounts, workflows and situations it covers, and previews the past conversations it would have reached. Once activated, the correction is added to the agent only when that situation arises, and all corrections live in one place instead of being scattered through prompts.

Measured against a holdout

Each correction is compared with and without it using held-out sessions, so teams can see whether it helped. Mirrors positions itself beside observability tools, which show what happened, and evals, which test before shipping, rather than replacing either.

An SDK that stays out of the way

Teams install the SDK, send their agent's conversations and request corrections once per user turn. The application keeps its own model calls and tools. The SDK waits up to 1,000 milliseconds per turn by default, configurable, and if Mirrors does not answer in time the agent simply carries on without corrections. The site lists compatibility with LangChain, LangGraph, LlamaIndex, CrewAI, AutoGen, Pydantic AI, Google ADK, Mastra, the AI SDK, OpenAI, Anthropic, MCP, Claude Code, Codex, Cursor and Windsurf, among others.

Research roots behind the SDK

Aryan Singhal, the CEO, published six AI and machine learning research papers while in high school, including at NeurIPS and EMNLP, and later did ML research at UT Austin and UCSB. Aditya, the chief technology officer, built computer vision and agent workflows at several startups handling thousands of sessions a week at sub-second latency.

Key facts

Maker: Mirrors, Inc. Launched: 30 October 2026 on Y Combinator's launch board Founders: Aryan Singhal (CEO), Aditya (CTO) YC batch: Fall 2026 Pricing: not published; request access or book a demo Integrations: any agent framework via SDK; see list above Website: mirrors.dev