Home›Companies›Mirrors
Mirrors

The guidance system for production AI agents

Mirrors gives developers the infrastructure to control what production AI agents are told, and when. Today that lives in long prompts and skills the agent may never open, so agents keep breaking rules they've already been given. With Mirrors, a developer writes a rule once and defines where it applies, and Mirrors delivers it at the exactly where the agent needs to see it and measures whether it helped. It works with any agent framework and any model.
Active Founders
Aryan Singhal
Aryan Singhal
Founder/CEO
Co-Founder & CEO of Mirrors (YC F26). Published 4 papers at NeurIPS and EMNLP on how LLMs fail (3 as first author). Previously an ML researcher at UT Austin and UCSB. Studied CS/AI at Purdue.
Aditya Gupta
Aditya Gupta
Founder/CTO
Founder at Mirrors (YC F26). Studied Computer Engineering at Purdue.
Company Launches
Mirrors: Fix recurring AI agent mistakes with real-time guidance
See original launch post

Hey YC! We’re Aryan and Aditya, the founders of Mirrors. Mirrors helps developers maintain production AI agents by turning recurring mistakes into targeted corrections they can preview, apply, and measure.

https://youtu.be/pC8PkI86RGE

Your system prompt is becoming a history of everything your agent got wrong.

Your support agent repeats troubleshooting steps the customer already tried. You add a line to the prompt. It keeps pushing a retention offer after the customer declines. You add another.

After enough fixes, the agent is carrying instructions for dozens of situations, even when it’s doing something simple.

As those corrections spread across prompts and skills, there’s more to maintain: where each instruction belongs, which accounts it applies to, and whether it actually helps.

uploaded image

That’s the agent in our launch video: covered in sticky notes, trying to follow them all.

How Mirrors works

Mirrors surfaces recurring mistakes across your production runs. Say your support agent keeps restarting troubleshooting instead of building on what the customer has already tried. In Mirrors, you write a correction:

Before suggesting troubleshooting, check which steps the customer has already completed. Acknowledge the results and continue from there.

Choose the situations where it applies, then preview which past conversations it would have reached. Once activated, Mirrors adds it to your agent only in those situations. Afterward, compare outcomes with and without the correction to see whether it helped.

uploaded image

Corrections live in one place, scoped to the right accounts and workflows. When requirements change, you update one correction instead of juggling versions of your entire prompt.

uploaded image

Mirrors connects through an SDK and works with any agent framework. Your application keeps control of its model calls and tools.

Why we're building this

One of our design partner's agent hit 75 errors editing a single travel guide. The rule telling it to stop was already in its prompt four times. More instructions weren't the fix. Getting the right one to the agent at the right moment was.

We're building Mirrors so improving your agent doesn't leave you with an ever-growing collection of prompt patches to maintain.

The team

  • Aryan (CEO): Published 6 papers at NeurIPS and EMNLP on how LLMs fail (3 as first author). Previously an ML researcher at UT Austin and UCSB.
  • Aditya (CTO): Built computer vision and agent workflows at multiple startups serving thousands of sessions per week at sub-second latency.

Our ask

Tired of playing whack-a-mole with your agent? Bring us one failed run.

Book a time with us or email devs@runmirrors.com. We’ll work through what went wrong and how Mirrors can help.

mirrors.dev

YC Photos
Mirrors
Founded:2026
Batch:Fall 2026
Team Size:2
Status:
Active
Location:San Francisco
Primary Partner:Andrew Miklas