Operationsยท6 min read
Controlling Agent Costs Before They Control You
Agents fail open on spend. The handful of habits that keep a useful agent from becoming an expensive one.
Controlling Agent Costs Before They Control You
A chatbot costs one call. An agent loops โ and every loop spends tokens. Cost discipline is an operational skill, not an afterthought.
Where the money goes
- Long context re-sent on every turn.
- Over-long loops that retry instead of stopping.
- Frontier models doing clerical work.
Habits that pay
- Cap the loop. A hard limit on steps prevents a confused agent from spending forever.
- Trim context. Summarize old turns instead of resending them verbatim.
- Cache the stable stuff. System prompts and reference docs that don't change should be cached, not re-billed.
- Route by difficulty. Send the easy 80% to a cheaper model.
Make spend visible
Log tokens per run and watch the trend. Costs rarely spike โ they drift. The team that reviews a weekly cost-per-task number catches the drift while it's still cheap to fix.