Agent Cost Control

Where the money actually goes

Measured on a 2-page paper produced end to end by an agent, on the cheapest available model:

Almost none of that is the model writing the paper. It is the conversation history being re-sent on every request. A single writing phase averaged 1.7 million input tokens per agent call.

The same shape on a mid-tier model costs roughly $37, and on a frontier model considerably more. If a pipeline feels surprisingly expensive, this is why, and the fix is not a cheaper model.

The negative result: compacting harder costs more

The obvious optimisation is to compact conversation history aggressively so each request carries less. It was measured on the same task, against a baseline of 12.5M input tokens and $13.19:

Compacting harder was two and a half times more expensive. Sessions in the tight configuration reached 463 and 408 agent actions with 37 and 92 compactions. Starved of its working context, the agent re-read files it had already read, re-ran commands, and re-planned work it had already done. Per-request size fell. Turn count exploded. Total spend more than doubled.

An offline replay of the event stream had predicted the tight budget would cost 39 percent of the generous one. The prediction was worthless because it assumed the event stream stays the same when the budget changes. It does not. The budget changes what the agent does, and that feedback dominates everything else.

Treat a context budget as a safety ceiling against a pathological history, never as a cost lever.

What actually helps

Fewer turns, not smaller turns. Cost scales with how many times the agent goes around the loop. Give it what it needs up front rather than making it discover the workspace one tool call at a time.

Put the key files in the task. If the agent will obviously need results/exp1.json and the venue template, say so and let it read them once, deliberately.

Bound exploration. "Read the three files under results/ and write the Experiments section" is cheaper and better than "figure out what experiments were run".

Split long phases. A phase that will take 200 turns is better as three phases with checkpoints on disk between them. State on disk is free. Conversation history is not.

Watch the ratio, not the total. An input-to-output ratio above roughly 50:1 means you are paying to re-read rather than to think.

Before starting a long run

• Estimated cost stated to the user up front
• A ceiling set, and behaviour defined for reaching it
• Long phases split, with results written to disk between them
• Spend visible while running, not only at the end