Published on arXiv, CoCoDA addresses a scaling failure in tool-augmented language models. As tool libraries grow, retrieval costs balloon and prompt budgets break. The framework structures tools as nodes in a directed acyclic graph, with typed signatures, behavioral specs, and worked examples attached to each node.

Existing approaches treat tools as flat or text-indexed memory. That means prompt cost grows with library size. CoCoDA's Typed DAG Retrieval prunes by symbolic signature unification first, then ranks, filters, and disambiguates on progressively smaller candidate sets. The authors provide theoretical results showing sublinear retrieval time and monotone co-evolution under conservative updates.

The operating consequence is separating library growth from context cost. Successful trajectories fold into validated composite tools, and a DAG-induced reward credits composites by their primitive expansion size. Operators running small models with large skill sets should watch whether this pattern holds outside math and tabular reasoning benchmarks.