Designing Agent Tool Loops Around Per-Call Cost, Not Frontier Assumptions
Most agent bills come from the loop, not the model. This short runs the arithmetic from the post: a ten iteration loop sends 69,100 input tokens against 1,800 output tokens, because every iteration re-sends the system prompt, the tool schemas and every tool result so far. Full write-up: https://hackyjs.com/posts/designing-agent-tool-loops-around-per-call-cost-instead-of-frontier-model-assumptions What's covered: - the four input token buckets, and why tool schemas are a fixed per-call tax - a per-call meter that logs input, cached and output tokens plus wall time - splitting the planner from the executor, with loop caps and escalation criteria - why per-call timeouts don't compose: 9.6 percent at ten calls, 55 percent at eighty Note: the post treats Muse Ember's parameter count, context window, pricing and benchmarks as unknown. The numbers here are loop arithmetic you can rerun against any candidate model, not model benchmarks.