AI Agents Waste Tokens Re-Reading Unused Tools on Every Turn, Microsoft Offers Fix

AI agents that use large tool catalogs silently waste input tokens by resending the full JSON schema for every tool on every request turn, regardless of whether those tools are used. A 200-tool catalog at 250 tokens per schema can consume around 600,000 input tokens on tool definitions alone across a 12-turn task, dwarfing the actual conversation cost. Beyond cost, large catalogs with near-duplicate tool names degrade the model's tool-selection accuracy, not due to model limitations but due to genuine ambiguity in the options presented. Microsoft introduced a Tool Search feature at Build 2026 for Azure AI Foundry, replacing the flat tool list with two meta-tools that let agents retrieve a small ranked set of relevant tools on demand, keeping token usage roughly flat regardless of catalog size. The approach is most beneficial above roughly 30 tools, and can be further optimised by pinning frequently used or critical tools to bypass retrieval latency entirely.