Who you are
For the finance controller
You are accountable for what this spends. You need to know who commits money, what stops a run, and whether the numbers can be defended.
| Who this is for | You own the budget for agent work — as a finance controller, a CFO, or the engineering manager who signs for it. |
| What you need | Read access to the record. Setting a policy for what may be committed is a separate act, and a human one. |
| What you will not find | A price list. What this product controls is your spend with your vendors: it records what was committed, by whom, against which expectation, and it stops a run that exceeds its ceiling. |
Your first task: read how a run's ceiling is derived from its quote, because that one relationship is what actually stops spend. Everything else on this page assumes the distinction.
The three questions you will be asked, and where each is answered
| Go here | Because |
|---|---|
| Budgeting a run: the quote and the cap | "Who commits money, and what stops it?" Two numbers ride on every dispatch and they are different kinds of thing: a quote is a person's prediction and changes nothing; a cap is the ceiling the run is actually stopped at. Read this one first — everything else assumes the distinction. |
| The cost pricing registry | "Where do the rates come from?" The engine normalises every run onto one currency column and never invents a rate: for each backend the rate — or the declaration that there is no marginal rate at all — is written by a human into one file. No verb creates it, edits it, or guesses. |
| Tracing one change backward | "Can I defend a number?" The walk from a merged change back to the decision, the order and the cost attached to it. A cost figure nobody can trace to a decision is an anecdote. |
What actually controls spend, in the order the money moves
- Someone quotes the work before it starts. A person writes the expected cost onto the dispatch. It is a prediction, deliberately powerless — its only job is to be compared against the outcome afterwards.
- The ceiling is derived from that quote, not chosen separately — so a cheap quote cannot carry an expensive licence, and the relationship between the two numbers is a rule rather than a habit. How the cap is derived.
- The run is stopped at the ceiling, not warned. What happens at the ceiling, and the three distinct ways a run can disagree with its quote, are on the same page.
- Each disagreement is written down and read back — the register of what work of a given shape actually cost is what makes the next quote better than the last. This is the part that turns budgeting from a flag into a practice.
- Review is a cost decision too. How many independent reviews a class of work must pass — one, or two by models of different families — is a policy you set, and it is one of the larger levers on total spend. Review policy.
Two structural facts worth knowing before you plan
| Spend is per seat, and a seat carries a vendor login of its own | Capacity, limits and exhaustion are properties of a seat rather than of the fleet, which is why "we still have budget" and "this run can proceed" are different questions. ⚠ Several seats can be signed in to the same vendor account, so per seat and per account are not the same number, and a usage reading taken from one seat is not a reading for the others. Seats |
| Backends are not interchangeable in cost | Several vendors can do the same work at very different prices, and the product records the outcome per backend rather than assuming one. A supported list is here; which to use for which class of work is a decision you own. |
The reporting and economics material
All of it is published, in Cost management:
| Go here | Because |
|---|---|
| What a ledger row knows | The fields on one row, and which of them a report can be built from. |
| The dimensions of cost | The axes spend can be reported along, and which combinations the record supports. |
| One yardstick across vendors | How work priced in different vendors' units is compared on one column. |
| Matching the model to the task | Where a cheaper model is the right answer and where it is a false economy. |
| Cheap by planning | The largest lever on total spend, and it is not the model choice. |
The limits on a number you may be asked to defend
- A cost figure nobody can trace to a decision is an anecdote. The walk that makes it defensible is tracing one change backward.
- A rate is written by a human, never invented by the product. Where no marginal rate exists, the registry says so rather than guessing (the pricing registry).
- Each of those pages closes with what it does not cover. Read that row before you build a report on the page above it.