Field notes · running agents
What AI coding agents actually cost per task
Roughly ten cents for a small change, across two model calls. The interesting part is not the number — it is that the model bill turned out to be the cheapest thing about running agents.
Where the money goes
One task — take a goal in a sentence, work in an isolated checkout of a real repository, produce a reviewed change — costs about ten cents at frontier model pricing. The whole pipeline makes two model calls:
Five of seven stages involve no model at all. That ratio is the design: a language model is expensive judgement and a poor file-editor, so it decides and ordinary code does the work.
Why per-task beats per-seat for reasoning
Ten cents a task reframes the question. At that price the constraint on how much agent work you do is not the bill — it is how much output a human can meaningfully review. Which means the scarce resource in an agent system is attention, and the pipeline should be designed to spend the cheap thing to protect the expensive one.
Concretely: it is worth burning extra model calls to avoid one wasted human review. Re-planning rather than presenting a doubtful change, verifying before parking at the gate, refusing rather than guessing — each of those trades cents for minutes, in the right direction.
The line item nobody tracks: re-runs
Our worst case was eight attempts on one task before it produced a real diff. In tokens that is eight times ten cents — trivial. In every other respect it was expensive: each attempt needed reading, each failure needed diagnosis, and two of them parked confident summaries at the approval gate with empty patches attached.
None of that appears in a cost dashboard, because every run reported success. If you count cost per task rather than per run, and count a success with no result as a failure, the number stops lying.
What the gate costs, honestly
An approval gate costs seconds of human time per consequential action, and it is the only reason one of our runs did not ship a change that would have broken authentication for every user of a live product.
That asymmetry is awkward to put in a spreadsheet. The cost is continuous and visible; the benefit is occasional and invisible, because it takes the form of an incident that did not happen. Which is precisely why gates get removed in the quarter before they were needed.
Rules of thumb
Model calls are for judgement. If code can do a step deterministically, code should — it is cheaper, faster and it does not hallucinate a file path.
Count tasks, not runs. A task that took eight runs cost eight runs, and the reported success rate is not the completion rate.
Buy certainty with tokens. Extra calls that prevent a wasted human review are the best-value spend in the system.
Watch time-to-decision. The most expensive state an agent can be in is finished, correct, and waiting for someone to look.
FAQ
- How much does an AI coding agent cost per task?
- In our pipeline, roughly ten cents for a small, well-scoped change — two model calls at frontier pricing, plus sandbox compute that rounds to nothing. That covers planning and the change itself; the mechanical work of editing files, running tests and producing a patch involves no model calls at all.
- Why only two model calls per task?
- Because most of the pipeline is deterministic. One call chooses which files to read; a second produces the change plan. Everything after that — applying edits, running the test command, building the patch, evaluating policy, pushing a branch, opening a pull request — is ordinary code. Model calls are for judgement, not for work a program does better.
- What is the real cost of running coding agents?
- Human attention and re-runs, not tokens. A task that needs three attempts costs triple in tokens and vastly more in review; an approval that sits unread for hours costs more than the run that produced it. Optimising the token bill while ignoring re-run rate and time-to-decision is optimising the cheapest line.
- Does a human approval gate make agents more expensive?
- It adds seconds of human time per consequential action and removes the cost of shipping a wrong change — which in one of our runs would have broken authentication for every user. Measured against a single incident, the gate is not a cost centre. Measured against nothing going wrong, it looks like pure overhead, which is why it is usually removed just before it was needed.
- How do you keep agent costs from growing quietly?
- Track cost per task alongside re-run rate, and treat a run that reports success with no result as a failed run when you count. Otherwise the cheap number stays cheap on paper while the same task is quietly attempted eight times — which is exactly what one of ours did before producing a real diff.
see the whole flow, live
Everything on this page — the queue, the evidence, the hold-to-approve, the audit trail — is working in the VIVUU terminal as an interactive preview on realistic mock data. No sign-up.