HyperTool: Beyond Step-Wise Tool Calls for Tool-Augmented Agents

HyperTool introduces a unified executable interface for tool-augmented LLM agents, addressing execution-granularity mismatches by allowing models to invoke tools with code blocks that manage intermediate results internally. This approach enhances the efficiency of multi-step tool workflows, as demonstrated by significant accuracy improvements on the MCP-Universe benchmark: from 15.69% to 35.29% for Qwen3-32B and from 9.93% to 33.33% for Qwen3-8B, outperforming GPT-OSS and Kimi-k2.5. This development is crucial for practitioners as it streamlines tool integration, reducing context consumption and improving overall model performance in complex task environments.