You have 40 tools registered on an agent. What happens?

Tool-selection accuracy degrades, and it degrades faster than most people expect. Redundant tools with overlapping descriptions actively confuse routing — the model has to disambiguate between things that look similar in description space, and that’s exactly where it starts picking the wrong one or hedging with an unnecessary extra call.

The fixes, roughly in order of how often I reach for them:

1. Consolidate overlapping tools rather than registering near-duplicates.
2. Write tool descriptions around when to use this tool, not just what it technically does — that’s the actual disambiguating signal.
3. For genuinely large tool sets, don’t present all of them at once — retrieve a relevant subset per request, turning tool selection into a retrieval problem rather than a 40-way classification the model has to do cold on every call.