Progressive MCP Tool Routing: How We Cut Agent Hallucination by 40% in a 47-Tool Environment

Description and use

Progressive MCP tool routing cut hallucinations by 40 % and token usage by 65 % in a 47‑tool environment, boosting first‑call success to 89 %.

Problem

AI agents with large toolsets (47 MCP tools) suffered from cognitive overload, spending 70% of reasoning on parsing tool definitions, leading to 22% hallucination rate and high token consumption.

Outcome

Progressive MCP tool routing reduced hallucinations by 40%, token consumption by 65%, and increased first-call success from 68% to 89%, while maintaining scalability as tool count grew.

Workflow

• 1. Agent receives high-level task and accesses a compact 3,000-token router guide describing 5 tool categories.
• 2. Agent selects a category (e.g., Data Analysis).
• 3. System performs semantic search within the selected category using vector embeddings to retrieve top 3 relevant tool schemas.
• 4. Agent selects a specific tool from the retrieved list.
• 5. System injects the full detailed schema of the chosen tool into the context for execution.

Evidence

Agents spent 70% of their reasoning budget parsing tool definitions instead of executing tasks.

Source

Hallucination Reduction: Incorrect tool calls dropped from 22% to 13.2%, a 40% relative reduction.

Source

Token Efficiency: Average token consumption per task plummeted from 52,400 to 18,340 tokens, a 65% savings.

Source