Point it at an existing agent, a stream of unlabeled production traces, and a small labeled holdout set.
An LLM judge scores unlabeled production traces as they stream.
A proposer reads failed traces and writes one targeted harness update at a time, such as changes to prompts, hooks, tools, or subagents. The update is kept only if it improves holdout accuracy.
On tau-bench v3 airline, meta-agent improved holdout accuracy from 67% to 87%.
We open-sourced meta-agent. It currently supports Claude Agent SDK, with more frameworks coming soon.
Try it here: https://github.com/canvas-org/meta-agent