Tool Schema Pruning Protocol (Function Calling Minification)

Overview

When configuring autonomous agents with Tool / Function Calling capabilities (OpenAI tools: [...] or Anthropic tools: [...]), frameworks inject complete JSON Schema definitions for every available tool into the input context.

In systems with 15 to 25 tools, bloated JSON Schemas with verbose parameter descriptions, empty example arrays, and redundant titles consume 4,000 to 8,000 input tokens on every single turn. This "Schema Tax" is paid on every user message and tool result, accounting for up to 50% of the entire session token bill.

The Tool Schema Pruning Protocol eliminates decorative metadata, strips self-evident descriptions, and dynamically exposes only the tools relevant to the active operational phase.

Bloated JSON Schema vs. Pruned High-Density Schema