Definition
Plain language
A map of which AI tools can plausibly feed their outputs into which other tools.
As stated in the literature
A directed graph over a tool population with edges scored by LLM judgments about whether one tool's output schema can plausibly serve as another's input; used in Firefly to guide multi-tool trajectory exploration.
Why it matters: It turns the combinatorial explosion of possible tool sequences into a structured map, so training data generators focus only on chains that could plausibly work.
For example, the graph might have an edge from a 'search flights' tool to a 'book flight' tool because the first's output schema fits the second's input.
Heard on the show
“The piece that does the heavy lifting here is what the authors call a tool compatibility graph, and I think the cleanest way to picture it is as a recipe book.”Episode 059 — Firefly's Inversion: Building Verified Tool-Call Training Data by Working Backward