When software agents use tools to complete tasks, they often need to chain multiple tool calls together, where the output of one tool becomes the input for the next. The dependencies between tools are difficult for language models to infer just from tool descriptions, especially in large enterprise tool catalogs.
This research builds a graph of these dependencies once, offline. Each node is a tool, and edges show which tools can supply output to others. The authors then use this graph to score possible next actions by estimating how valuable they will be several steps ahead, given a final goal. Three different scoring methods were tested: one based on exact graph calculations, one using a learned graph neural network, and one using a language model that also sees the current tool arguments.
Only the language-model-based scorer improved performance, beating a standard graph-retrieval method by an average of 8.3 points across all tested scenarios. The purely structural scorers underperformed because they could not detect when a tool was called with wrong arguments. This shows that for multi-step tool use, the graph's value comes from combining structural dependencies with runtime state information.
