Definition
Plain language
A system where one AI watches another AI run an evolutionary search and rewrites the rules of the search as it goes.
As stated in the literature
An agentic evolution framework with a meta-agent that edits the underlying search mechanism (selection rules, prompts, notes) between segments, enforced by a sandboxed harness that prevents reward hacking.
Why it matters: Most evolutionary search is brittle to fixed hyperparameters, and letting a meta-agent edit the search itself can adapt to whatever the problem actually needs.
For example, after a poor batch of results, AEvo's meta-agent might rewrite the selection rule to favor more diverse candidates and resume the search.
Heard on the show
“… The authors run AEvo on the task and end up at 1138 cycles within 100 iterations — which the authors say is, to their knowledge, …”Episode 046 — When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall