Hey there,
My top two priorities for OpenClaw are, obviously, getting the job done, but also doing it while keeping token usage to a strict minimum so I don't blow my LLM budget.
By using a local model and digging into the logs, I realized that the sheer number of skills and tools available to the agent was massively bloating my request context. After feeding those logs into Gemini, I spotted a bunch of unnecessary tools and skills that had nothing to do with my specific task.
That got me thinking: it would be way more efficient to create task-specific agents by hand-picking only the tools they actually need. For some tasks, I managed to cut the token count by almost 10x!
I’d love to know—what are your main strategies for optimizing token consumption?