Definition
Plain language
Pulling an AI agent's recurring thinking moves out of its old transcripts and turning them into reusable tools.
As stated in the literature
The process of mining successful agent reasoning traces, labeling recurring reasoning moves, and synthesizing them into named pseudo-tools; argued to surface latent reasoning structure the source agent generated inconsistently rather than merely distilling its behavior.
Why it matters: It turns an agent's inconsistent but useful reasoning habits into dependable tools, surfacing structure the agent only sometimes used on its own.
For example, after noticing an agent often breaks a problem into 'check the edge cases first,' that recurring move is packaged into a named, reusable tool.
Heard on the show
“… And there's a line in the paper that nails the thesis — that trace induction isn't merely distillation; it can surface reasoning structure the source agent generated inconsistently …”Episode 110 — How an Agent Got 44 Points Better by Mining Its Own Scratch Paper