Index your tools by input/output schema as a hypergraph so the agent retrieves tools by what it's missing, not by keyword similarity

arXiv 2608.02650·low signalXBlueskyLinkedInCopy link

HyperAgent builds a directed Tool-Schema Hypergraph where each tool is a hyperedge running from its required input-schema nodes to its output-schema nodes, then uses deficit-oriented expansion at runtime: it identifies which schema requirements are currently unmet and retrieves only tools that can supply them. The retrieved task-relevant subgraph guides construction of a schema-aware Task DAG, replacing textual inference over a flat tool list. On AppWorld it improves task completion while simultaneously reducing redundant API calls, LLM interactions, and token consumption — the reported improvements are directional rather than tabulated in the abstract, so treat the magnitude as unverified, but the indexing pattern is directly applicable to any agent with an oversized MCP tool surface.