Meta Agent Search turns prompts, tools, and control flow into code that a meta agent can generate, run, score, and place in an archive; the reported designs beat hand-built baselines and transfer across domains and underlying models. This gives Compositional Intelligence a mechanism for discovering runtime organizations rather than prescribing one ideal agent. The archive makes successful programs reusable material for the next round of search.
Agents are defined as programs in a Turing-complete language. That choice sets the search space: prompts, tool use and control flow are all just code, so anything expressible is reachable.
A meta agent writes new agent programs, runs them, scores them, and adds what it finds to a growing archive it can read from. The authors call this Meta Agent Search.
Discovered designs beat hand-built baselines, and they transfer: agents found in one domain hold up in another, and agents found with one underlying model hold up on a different one.