The Menu Is an Execution Prior: State-Path Tool Menus for Online Agents

The State-Path Tool Menu: A major shift for Online Agents

for online agents, one of the biggest challenges is handle the vast array of tools and interfaces at their disposal. According to editorial research analyzed by Groundwork, With thousands of tools to choose from, agents can easily get overwhelmed and struggle to complete complex tasks. In this article, we'll explore the concept of the State-Path Tool Menu, a revolutionary approach to tool menu design that's changing the game for online agents.

The State-Path Tool Menu is a short, ordered subset of available tools shown to an agent before execution, enabling them to call only tools in this menu.

What's the Problem with Current Tool Menu Design?

Current constructors rank tools by request relevance, which can surface the final action while omitting or delaying less obvious producers. This can lead to a range of problems, including:

• Incomplete tool menus: Current menu design often fails to include all the necessary tools for a task, leading to incomplete chains and frustrated agents.
• Delayed execution: By omitting or delaying less obvious producers, current menu design can cause agents to spend more time searching for the right tools, slowing down execution.
• Inefficient tool usage: Current menu design can lead to agents using the wrong tools or using them in the wrong order, resulting in wasted resources and decreased efficiency.

Introducing the State-Path Tool Menu

The State-Path Tool Menu is a novel approach to tool menu design that addresses these problems head-on. By learning the state path, a pre-execution route from the observable request state to the desired outcome, the State-Path Tool Menu can provide a complete and efficient tool menu for agents.

How Does the State-Path Tool Menu Work?

The State-Path Tool Menu consists of three key components:

1. Encoder: The encoder represents which tools can run from the current state, how their outputs satisfy later inputs, and which orders recur in training paths.
2. Retriever: The retriever covers an executable entry, the missing-input producers, and the final action.
3. Reranker: The reranker places producers before consumers, ensuring that the most important tools are listed first.

Experimental Results

Our experiments on ToolBench show that the State-Path Tool Menu raises online success from 0.737 to 0.898, outperforming retrieval, reranking, generation, and routing baselines without changing the agent. The State-Path menu also covers more complete chains with 32 tools than the official list covers with 128, and its success gain persists across executor families with different model capacities.