Definition
Plain language
A tool an AI can call by name, but whose 'body' is just another model improvising the answer rather than real code.
As stated in the literature
A named, typed callable with no deterministic implementation; invoking it bundles the description and inputs into a fresh LLM prompt that produces the result, bridging callable interfaces and judgment-laden reasoning that resists clean code.
Also called: pseudo-tools
Why it matters: It matters because it lets agents package fuzzy, judgment-heavy reasoning behind a clean tool interface that ordinary code can't capture.
For example, an agent might call a 'judge_argument_strength' tool that, instead of running real code, simply prompts another model to weigh in.
Heard on the show
“The authors call these pseudo-tools.”Episode 110 — How an Agent Got 44 Points Better by Mining Its Own Scratch Paper