Abhibhaw Asthana

Building now / AI agents · Reliability / In development

Peira: AI agents that improve from their own failures

AI agents keep failing in production, and teams find and fix each failure by hand. Peira finds repeated failures in an agent's traces, proposes a fix, and proves it before a person approves it.

User
Teams that run AI agents in production and find out about failures from users.
Problem
Agent failures repeat quietly. Finding them means reading traces by hand, and testing a fix means hoping it does not break something else.
Decision
Find repeated failures in production traces. Let a coding agent propose a fix. Prove the fix by replaying past conversations against a world rebuilt from the traces. A person approves every change.
Not built
Automatic deploys of fixes. Every change arrives as one pull request with evidence, and a person decides.
Status
In development. No customers yet.

The problem

An agent that works in a demo fails in production in small, repeated ways. Teams learn about it from users, read traces by hand, and hope a fix does not break another case.

How it works

Find and fix
  1. Production tracesLangfuse, LangSmith, OpenTelemetry
  2. grouped into
  3. Repeated failures
  4. coding agent proposes
  5. Fixprompt, tool or code
  6. replayed on
  7. Past conversationsin a world rebuilt from traces
Ship
  1. Pull request with evidence
  2. a person approves
  3. Fix ships
  4. joins
  5. Protected setlater changes cannot make these worse
Each fix is proven on past conversations, and a person approves it.

The replay rebuilds the apps and data the agent used from its traces, so it never touches real systems.