Why Token Budgets Fail With Background Agents

Quick answer

Token budgets fail with background agents when they cover the whole job instead of the current phase. A long-running agent needs a budget, a progress signal, and a stop condition for each attempt.

The whole-job budget problem

A $20 budget sounds safe until the agent spends $18 discovering that the first step is blocked. The cap worked, but the workflow did not.

Background work should be split into phases: research, implementation, validation, deploy, and closeout. Each phase gets its own expected output and budget.

Retries need their own accounting

A retry is not free context. If attempt two loads the same files, repeats the same search, and re-runs the same failed build, the spend doubles while progress stays flat.

Track attempts separately so the second attempt has to justify itself.

Progress signals matter

The cheapest agent is not the one that uses the fewest tokens. It is the one that stops when it has no new evidence.

Require visible progress: files changed, tests run, commits made, artifacts produced, or a clear blocker recorded.

How Clawback helps

Clawback treats spend as workflow evidence. It helps answer whether tokens became progress, or whether a background agent just looped quietly.

Estimate agent task costs

FAQ

Is a token cap still useful?

Yes, as a last-resort guardrail. It should not be the only control.

What should each phase record?

Expected artifact, current state, token spend, model, attempt ID, checks run, and next validation step.

When should an agent stop?

When the same blocker repeats without new evidence, or when the phase budget is no longer likely to produce the required artifact.