Adding tools feels like progress. Until the agent starts slowing down and missing obvious instructions.
Every tool definition lives in the prompt budget. In a crowded context window, that cost is real.
The silent tax
People count tool capability, not tool overhead. I count both. A tool that gets used once a week should not sit in every session.
Unused tools are just token rent. You pay it on every turn.
What too many tools break
- Selection quality drops. The model picks the wrong tool more often.
- Latency rises. Bigger prompts mean slower calls.
- Instruction focus drops in the middle of the window.
- Debugging gets harder because failure paths multiply.
This is the same pattern as bloated microservice APIs. More surface, more mistakes.
My trim policy
I keep only tools tied to current work. Everything else stays disabled. If a tool cannot prove weekly value, it gets removed.
Then I wrap core actions in stable CLI commands with clear IO. Better tool use, less improvisation.
Small set, strong checks
A lean tool set plus strong back pressure beats a giant toolbox every time. Fewer choices, better execution.
Agents are not blocked by missing tools. They are blocked by unclear systems.