More tools are not better tools. When you hand a large language model a sprawling surface of 120+ functions, its reasoning gets worse, not better. A curated, well-named set wins.
Why a huge tool surface hurts
- Token cost. Every tool definition is loaded into context. A giant surface eats the budget the model needs for actual reasoning.
- Choice paralysis. Faced with near-duplicate tools, models pick the wrong one or chain calls inefficiently.
- Maintenance drag. Each tool is a contract to keep current as the underlying API shifts.
What a curated surface looks like
Instead of mapping every API endpoint one-to-one, group the work the way a marketer thinks: “summarize performance,” “propose a budget change,” “find fatiguing creative.” A few dozen intent-shaped tools cover the real jobs and keep the model's reasoning sharp.
The goal is not maximum coverage of an API. It is maximum clarity for the model doing the work.
This is why AdNexus favors a focused tool surface over a raw firehose: the agent reasons better, costs less to run, and proposes cleaner drafts.