HyperAgent teaches agents to schedule tool calls according to procedures

The problem with tool-use agents often appears to be a reasoning problem: the model was unable to determine which APIs to call and in what order. But there is a more mundane reason for this error—tools are usually described in text, and their actual connections between inputs and outputs remain implicit.

The HyperAgent approach proposes shifting planning to the data schema level. The authors represent tools as hyperedges in a directed Tool–Schema Hypergraph: input schemas become nodes, and the tool connects them to output schemas. This allows for describing not just a list of available APIs, but the dependencies between what the agent already has and what it can obtain with the next call.

For a specific task, the system first identifies a relevant fragment of the graph and constructs a schema-aware Task DAG—a directed acyclic graph of subtasks. During execution, the plan is not considered fixed: HyperAgent searches for unresolved requirements in the current state and extends the graph with tools that can resolve them.

This is an important distinction from the approach of “letting the LLM figure out the composition of tools based on descriptions”. The agent gains a more explicit mechanism for checking compatibility: if an input schema is missing for the next step, the search begins by addressing this deficiency rather than by exhaustively enumerating potentially suitable APIs.

In experiments on AppWorld, HyperAgent demonstrated higher task success rates compared to baseline agents, while simultaneously reducing the number of unnecessary API calls, interactions with the LLM, and token consumption. The paper does not provide specific improvement metrics, so this cannot yet be taken as proof of universal superiority.

The practical takeaway for agent system developers is simple: the tool catalog should be designed as more than just a set of text instructions for the model. Explicit input and output schemas can serve as a working layer for planning, especially when an error in a single call triggers a chain of futile attempts.

That said, HyperAgent does not eliminate the need for high-quality descriptions nor does it automatically solve all execution problems in the external environment. It formalizes specifically the part related to data dependencies and tool selection. The next question is how well such a scheme translates from AppWorld to unstable APIs, ambiguous data, and tools with side effects.

Source: cs.AI updates on arXiv.org