A correction with context.
Record the instruction with its scope: which customer, project, or workflow it belongs to.
Backed by Y CombinatorDon’t let the same agent mistake reach another customer.
You’ve already told your agent what went wrong. We’re building Mirrors to carry that correction into its next decision, even in a long-running task.
A founder will reply within 24 hours.
For teams putting agents into production Early design partners
You corrected it once.
Then it happened again.
“Get approval before sending a customer email.” It’s a simple correction. But later in the workflow, the agent is preparing another email. Will that instruction still shape its next decision?
For teams running agents in production, a forgotten correction can become another customer-facing mistake. We’re working on the gap between giving feedback and seeing it followed on the next task.
THE NEXT DECISION MATTERS
Bring the correction back.
When the agent needs it.
Here’s the workflow we’re building toward.
- 01 THE CORRECTION
“Get approval before sending a customer email.”
Scoped to this customer’s workflowCapture what should change, and where that instruction applies.
- 02 LATER IN THE TASKThe agent prepares
another customer email.This is the moment the earlier correction needs to be in context.
- 03 GUIDANCE TO SURFACERELEVANT CORRECTION
Ask for approval before sending.
Bring the instruction back before the next decision. Then inspect whether the agent followed it.
Prototype direction. This example illustrates the intended flow; it does not show an enforced approval gate.
MIRRORS / IN DEVELOPMENT
What we’re building.
A focused way to carry human corrections into an agent’s work.
Relevant guidance, brought back.
Bring the applicable correction into context as a long task progresses toward the next decision.
An output you can inspect.
Check the result against the correction, so your team can see what was followed and what still needs work.
The test is what happens next.
Did the right correction reach the agent? Did the output follow it? We will work with your team to evaluate those questions on a focused workflow, using the corrections and outcomes that matter to you.
EARLY DESIGN PARTNERS
Start with one
recurring mistake.
We’re looking for technical teams putting agents to work on real customer workflows, with a specific correction that keeps getting lost.
Bring us that example. We’ll talk through where it happens, what a better outcome looks like, and whether a focused prototype makes sense together.
STARTING POINT
One workflow. One known failure.
WORKING TOGETHER
Directly with the Mirrors founders.
SCOPE & TERMS
Discussed with your team.
A founder will reply within 24 hours.
FAQ
Common questions.
What kind of agent risk are you working on?
A mistake your team has already corrected, but the agent makes again: sending a customer email without approval, missing a project-specific requirement, or ignoring an instruction later in a long task. We’re building a way to capture that correction, bring it back when it’s relevant, and check the resulting output.
How is this different from adding to the system prompt?
A system prompt is still useful. The question we’re exploring is how to carry specific corrections across long workflows without putting every instruction in every context. Our prototype works toward scoping guidance to the right customer or project and surfacing it at the relevant decision.
Does Mirrors guarantee an agent will follow a correction?
No. This is an early prototype, and we have not established a measured reduction in mistakes. Our goal is to make prior corrections available at the relevant moment and make it possible to inspect whether the output followed them. Deterministic permission checks, approval gates, and other safeguards still matter.
How would Mirrors fit into our agent stack?
That is part of the design-partner conversation. We’ll look at where your team gives corrections, how your agent’s workflow is structured, and what you can inspect before an action happens. Integration scope and data requirements depend on that workflow; we are not offering a universal, self-serve integration today.
What does becoming a design partner involve?
Start by telling us what your team is building and one recurring agent mistake you want to address. A founder will reply within 24 hours. From there, we can discuss the workflow, a focused prototype, and how we would evaluate it together. Scope and any commercial terms are agreed directly with you.
Which agent mistake keeps coming back? Let’s start there.
Tell us about your workflow and the correction you need your agent to remember.
Talk to the foundersA founder will reply within 24 hours.