Definition
Plain language
An AI agent whose job is to watch another AI agent and rewrite the rules it's following.
As stated in the literature
In agentic-evolution frameworks, a coordinator agent that edits the underlying search mechanism — selection rules, prompts, persistent notes, skill files — between segments of an inner optimization loop.
Also called: meta-agents
Why it matters: It moves optimization from tuning the model to tuning the procedure around the model, which is often cheaper and easier to inspect.
For example, after watching its inner agent fail twice on similar errors, the meta-agent edits the prompt to add a 'check for this case first' instruction before the next attempt.
Heard on the show
“Right — and before we get into what the meta-agent actually does, it's worth saying that the cheating runs aren't dumb.”Episode 046 — When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall