Your agent's biggest token cost is not the conversation. It is the tools.
Context bloat due to tools needs to be talked about far more than it is. Every turn, an agent pays to declare its tools, select from them, and call them. At thousands of tools, the definitions alone run past a million tokens. It is an underestimated cost problem, and at scale it will bite organizations hard.
The research behind this is timely. It proposes a different approach entirely: stop listing tools to the model. Treat selection as recommendation. The agent asks for what a task needs and gets only that, when it needs it. Measured end to end: the bill dropped by up to 40%, and coverage of every tool a task actually needs rose 57% over the standard baseline.
Context bloat due to tools needs to be talked about far more than it is. Every turn, an agent pays to declare its tools, select from them, and call them. At thousands of tools, the definitions alone run past a million tokens. It is an underestimated cost problem, and at scale it will bite organizations hard.
The research behind this is timely. It proposes a different approach entirely: stop listing tools to the model. Treat selection as recommendation. The agent asks for what a task needs and gets only that, when it needs it. Measured end to end: the bill dropped by up to 40%, and coverage of every tool a task actually needs rose 57% over the standard baseline.
Missing one required tool can derail a multi-step agent task.View quoted post on X
@avoguru explains how agents can find the right tools among thousands, how to evaluate retrieval accuracy, and when a fixed list works better.
aaif.io/blog/how-to-gi…