Why It Costs What It Costs
Two agents doing the same job can cost wildly different amounts, and the driver is usually invisible on a demo: how many steps each takes and how much it carries in each step. Understand that and you can spot a runaway bill before it arrives.
An agent's cost is, roughly, the cost per step times the number of steps. Both sides bite. Per step: an agent that resends its entire history every turn gets more expensive as the task goes on. Number of steps: an agent that takes 14 steps to do what another does in 6 costs more even on a cheaper model. This is why "just use a cheaper model" often disappoints — if the agent is inefficient, step count dominates, and a cheap-per-step agent that wanders can cost more than an expensive-per-step one that finishes fast.
The practical leadership move is to ask for cost per completed task, not cost per model call — and to watch whether it's stable as tasks get harder. A team that can't tell you cost per completed task doesn't yet know its unit economics, which is exactly the number that decides whether an agent product makes money.
The cheaper model that cost more
Your team switched to a cheaper model to save money, but the monthly bill barely moved. What's the likely explanation, and what should they measure instead?
Cost is per-step times steps, and if the agent takes many steps — wandering, retrying, re-reading — then step count dominates and a cheaper per-step model barely dents the total. Have them measure cost per completed task and average steps, then attack the real driver: a tighter stopping rule and leaner context to cut steps and per-step size. Often the biggest savings come from making the agent finish in fewer steps, not from a cheaper model — an efficient agent on a pricier model can beat a wasteful one on a cheap model.
Loopcost: the bill hiding in the step count
Loopcost's agent worked, but the unit economics didn't: at scale it was losing money per task, and the team assumed the model was too expensive. When they finally measured, the culprit was steps — the average task took 18 steps, many of them redundant re-reads, and the agent resent its whole history each time, so late steps were huge. The model price was a minor factor; the loop's inefficiency was the bill.
They trimmed context each step (summarize old turns, drop stale results) and tightened the stopping rule to cut wandering; average steps fell to 7 and cost per completed task dropped by roughly 65% — on the same model. The founder's lesson: for agents, "what does it cost?" really means "how many steps, carrying how much, per finished task?" — and that's the number to watch before scale turns it into a crisis.
Tessera tracks cost per resolved request, not per model call, and notices early that a few request types balloon in steps. Trimming context and tightening the stopping rule keeps the number flat as volume grows — so the agent stays cheaper than the human hours it replaces.
Quiz · Chapter 6
- An agent's cost is roughly:
- "Just use a cheaper model" often disappoints because:
- The number a leader should ask for is:
- The biggest savings often come from: