The Problem: Tool Overload and Poor Prioritization

Modern AI agents often struggle when faced with massive libraries of available tools. Standard approaches typically rank tools based on simple request relevance. This often surfaces the final action tool while omitting or delaying the necessary prerequisite tools required to generate the inputs for that final action. This disconnect leads to failed execution chains, as the agent lacks the immediate context of how to prepare data for the final step.

The Solution: State-Path Tool Menus

The authors introduce the "State-Path Tool Menu," a framework that treats the menu of available tools as an execution prior. Rather than showing all tools, the system provides a short, ordered subset. The core innovation is the "state path"—a pre-execution route that maps the current observable request state to the desired outcome.

The framework functions through three components:

• Encoder: Represents which tools are executable from the current state, how their outputs satisfy downstream inputs, and identifies recurring successful orderings from training data.
• Retriever: Selects the specific executable entry points, the producers of missing inputs, and the final action tool.
• Reranker: Reorders the selected tools to ensure that prerequisite "producer" tools appear before the "consumer" tools that require their output.

Performance and Impact

Testing on the ToolBench benchmark demonstrated significant improvements, raising online success rates from 0.737 to 0.898. The State-Path approach is highly efficient: a menu of just 32 tools generated by this method outperformed the standard official list of 128 tools. Furthermore, these gains were consistent across different model architectures, suggesting that the improvement is rooted in better task structure rather than model capacity alone.