Pay for the step that needs a model.

Use the cheaper shape.

A known path is a workflow. The model works inside a step, and the next step starts when the last one finishes. An agent is for the step that has to choose. An agent on a fixed process spends the run deciding things a rule already knows.

Put the strong model on the hard step.

Extraction, a format check, and a classification do not need the same model as a judgment. The expensive call is reserved for the step the spec says is hard. The rest of the path stays small.

Write the spec before the run.

A prompt that has to rediscover the outcome is paid for on every retry. The spec says what will be true and what must not change. The model drafts against that. Rewrites are the quiet cost.

Send only the context the step needs.

Retrieval brings the page, the clause, or the diff. The whole repository, the whole policy, and the whole thread do not ride along on every call.

Stop the loop, and keep the repeat.

A limit on retries, tool calls, and length is a cost control as well as a guardrail. A run that keeps going after the goal is met is spend with no outcome. The same document, the same test, and the same policy answer are not paid for twice. A cheap run that a person rejects is the expensive one. Cost is counted per accepted outcome.

What a client gains.

• The model is used where a decision is required

  A rule carries the known path. The agent is reserved for the step that has to choose.

• The context is the part the step needs

  The page, the clause, or the diff. Not the whole system on every call.

• A run stops at the boundary

  Retries, tool calls, and length have a limit. The loop does not continue after the goal is met.

• Spend follows the outcome

  The count is an accepted result, not the length of the prompt.

This is how both offers stay honest. AI used to engineer the software, and AI engineered into the business, are commissioned against an outcome. The spend follows that outcome.