Image: Mirrors
Why it mattersTeams running agents in production keep adding rules to the system prompt and the prompt grows without end. Mirrors keeps each fix scoped and measurable, which is the only way the pile gets smaller.
Mirrors launched on Y Combinator's Fall 2026 batch page, with 12 votes at time of writing. The pitch is specific. Your agent's system prompt turns into "a history of everything your agent got wrong". Every production failure gets fixed with one more line. Mirrors says a team can keep the fixes without the prompt growing.
The company is Aryan Singhal and Aditya Gupta, both based in San Francisco. Singhal says he published six AI research papers in high school. Gupta built computer vision and agent workflows at earlier startups.
What the tool does
Mirrors runs on your agent's production conversations through an SDK. The company says the SDK waits at most 750 milliseconds per user turn, and the limit can be shorter. If Mirrors does not answer in that window, the agent continues without a correction.
The core idea is a correction: a short instruction with a scope. The founders use a support agent example on their launch page. A support agent keeps restarting troubleshooting instead of continuing from where the customer already got to. The team writes: "Before suggesting troubleshooting, check which steps the customer has already completed. Acknowledge the results and continue from there." Then they choose the situations where it should apply. Before turning it on, they preview which past conversations the correction would have reached. After turning it on, they can compare outcomes with and without it.
The founders say the alternatives both fall short in different ways. A line in the system prompt reaches every conversation, so unrelated runs still carry the rule. A skill is read only when the model decides it is relevant, so the fix can be missed. Mirrors keeps each correction in one place with its own scope, applied when its situation comes up.
The one real example and what it cost them
The launch page gives one case for why they built it. One of their design partners had an agent make 75 errors editing a single travel guide. The rule telling the agent to stop was already in the system prompt four times. "More instructions weren't the fix," the founders write. "Getting the right one to the agent at the right moment was."
That is one design partner and one number, both from the company. The launch does not name the partner. No outside source has measured how much Mirrors reduced the error rate. No outside source has tested the shape on a different kind of agent either. The 75 figure is what the company says. Nobody else has reproduced it.
The company says Mirrors works with any agent framework and does not take over the agent's model calls or tools. It works alongside evals and observability. Evals check against test cases before release. Observability shows what the agent did. Mirrors acts on the recurring pattern in production.
The question Mirrors is pointing at is one every team running agents in production hits sooner or later. The system prompt is the easiest place to put a fix. It is also the worst place to leave it. Every unrelated run carries every rule. A tool that scopes each correction to its situation is worth trying. The company is early. One design partner is not a track record.
Source
- Mirrors launch on Y Combinator, YC Fall 2026.
- mirrors.dev, company site.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
