For agent-driven demand, is the declaration a cap rather than a forecast?

[anacay]

A declaration is a signed statement of expected demand (Section 5.4). That assumes the declarer predicts something outside its control.

An agent fleet is different. Its owner holds the dial: budget caps, token ceilings per task, retry rules. A volume forecast there is a forecast of your own future choices.

So maybe the declaration should be the policy itself. The cap, the burn-rate alarm, and the condition for raising it. Score the team on cap discipline, not forecast error. Alibaba stated the limit case for flash sales years ago: capacity can be predicted but not planned, only limited. Version 7.9 of the seed paper adds this idea as a labeled exploration in Section 11.

The doubt: caps can be theater. If every team caps at triple its expectation, nothing improved.

Questions for the group:

• Who runs agentic or batch AI workloads under declared caps today? Do the caps bind?
• How do you tell user-driven AI demand (still a forecast) from machine-driven demand (a policy)?
• What belongs in the policy besides the cap and the raise condition?

All examples invented.