CoCoDA: Co-evolving Compositional DAG for Tool-Augmented Agents

A novel framework known as CoCoDA (Co-evolving Compositional DAG for Tool-Augmented Agents) has been introduced to tackle the issue of scaling tool libraries for language models. This framework integrates the planner and tool library within a unified code-native format: a compositional code directed acyclic graph (DAG). In this DAG, nodes signify either primitive or composite tools, while edges represent invocation dependencies. Each node contains a typed signature, description, specifications for pre/post-conditions, and examples. During inference, Typed DAG Retrieval optimizes candidate selection through symbolic signature unification, which minimizes prompt costs and facilitates efficient retrieval under a fixed context budget. This method differs from current approaches that view tools as flat or text-indexed memories, leading to increased prompt costs and obscuring the typed, compositional nature of executable code. The research is available on arXiv under identifier 2605.08399.