How are you keeping multi-step agent runs from blowing up token cost?
A three-step agent quietly turned into fifteen tool calls per task because each step re-sent the whole history. Trimming context to only what the next step needs and summarising older turns cut the bill by more than half. The practical additions: a short memory file the agent reads instead of the full transcript, a hard budget per run that stops and asks instead of burning tokens, and letting the cheap model handle routine steps.