The mandate is the product
An agent that asks about everything is a worse inbox. An agent that asks about nothing is a liability. The line between them is written down, and Steward never argues with it.
Who owns which change
| Situation | Who owns it |
|---|---|
| Workload is pinned | Nobody touches it. Steward reports and stops. |
Destination is not live | Refused. Production does not move onto a preview model. |
| Destination has less context than the workload needs | Refused. |
| Saving under $2.00 a month | Refused. Churn is not worth it. |
| Workload marked quality sensitive | Yours. Steward states the saving and waits. |
| Everything else | Steward's, and it does it. |
The workloads it looks after
| Workload | What it does | Context needed | Capability floor | Quality sensitive | Pinned |
|---|---|---|---|---|---|
| support-triage gpt-5.5 | Classifies inbound tickets and drafts a first reply. High volume, short answers. | 32,000 | light | no | no |
| contract-review gpt-5.5 | Reads long agreements and flags clauses. A wrong answer here is expensive. | 200,000 | flagship | yes | no |
| nightly-digest kimi-k3 | Summarises the day's activity into one page at 03:00. | 400,000 | light | no | no |
| code-review-bot glm-5.2 | Comments on pull requests before a human looks. | 128,000 | mid | no | no |
| board-pack gpt-5.5 | Writes the monthly board summary. Pinned by the finance lead. | 128,000 | flagship | yes | yes |
The capability floor is an absolute band, not a quantile of whatever is in the catalogue today, because a floor that moves with the candidate list is not a promise anyone can rely on.