Quality Is Part of the Agent Price

Coding agents make their consumption unusually visible. A run has a model cost, a duration, and an answer at the end. That invites a clean comparison: put the same task through several workers and choose the one that charges less. The arithmetic is real. The unit is wrong.

The business did not ask for an answer. It asked for a change that belongs in the product. Between those two things sit context assembly, retries, tests, review, integration, release, and whatever happens when the change meets production. A cheap run can push expensive uncertainty into every one of those stages.

That hidden transfer matters because the receiving system is usually a person. Someone reads the unfamiliar code, reconstructs the intent, repairs the edge case, or decides whether the proof means anything. The agent looks efficient because the human rescue is booked to payroll and disappears from the agent bill.

Two agents can finish the same task for different prices and still resist comparison. One may produce a patch that clears independent checks on the first route. The other may need repeated prompting, a second model, and a careful rewrite before it can advance. Price per attempt calls the second route cheaper. Price per accepted change may say the opposite.

Acceptance has to mean more than a green check the builder selected. The requirement must be specific enough to judge. The exact artifact has to face tests and policy outside the agent's control. High-consequence behavior needs evidence under conditions the implementation path cannot quietly soften after seeing the result.

Rejected work belongs in the denominator too. So do rollbacks and escaped defects. Excluding them rewards a route for manufacturing plausible output quickly and leaving the factory to discover later that it never produced completed work.

The scarce resource in an agent-run system is often not computation. It is the person who can tell whether an ambiguous result carries an acceptable consequence. If that person spends the day untangling routine patches, the factory has converted cheap generation into an expensive judgment queue.

Separate real judgment from repeatable repair. A product tradeoff, a new trust boundary, or an irreversible migration deserves an accountable human decision. Formatting corrections, known dependency rules, standard failure checks, and familiar recovery steps do not. Once the answer repeats, encode it in the route and stop purchasing the same decision from a person.

This is where jobs disappear. Review and coordination work that exists only to catch recurring machine mistakes will move into specifications, gates, and recovery machinery. The people who remain will decide what the system should do and what risks it may take. They will not be permanent shock absorbers for the cheapest model.

There is no reason to send every task through the most capable and expensive worker. A narrow, reversible change with deterministic proof should take a lean route. Ambiguous work that touches authority, money, identity, or durable data should get more reasoning, tighter permissions, and stronger verification. The task chooses the route through its consequences.

Make that policy mechanical. Classify the surfaces a task may touch, how easily its effects can be reversed, what evidence can prove it, and where a person must decide. Let less expensive workers compete inside those boundaries. Escalate when a known condition fires, not when somebody happens to notice that the output feels wrong.

Models and prices will keep changing. A stable routing contract lets the factory benefit without rebuilding its operating model around every release. The worker can change while the definition of acceptable work, the authority boundary, and the evidence required at the exit stay put.

Our position is simple: cost matters, but only after quality has been made operational. Count every attempt, every verifier, every human interruption, every recovery, and every failure that returns from production. Then divide by the work that earned its way through the whole route.

Use that record to change the factory. Move routine tasks to cheaper workers when they repeatedly clear the same bar. Strengthen context or proof when one class of work keeps failing. Stop a route that consumes more rescue than it saves. A measurement that never changes routing, authority, or controls is only an invoice with better labels.

The cheapest agent will win plenty of work. It should. But it has to win after the factory counts what the business actually receives, not when the model stops typing. Quality is not a premium added to the agent price. It is part of the price from the beginning.