Scheduled agents
Subscribe an agent to a verified feed, so it is told when something it cares about changes instead of asking on a timer.
Not built. Today an agent asks and Metera answers, one call at a time. Everything below is what a subscription would have to do to be worth shipping, written down so it can be held to it.
Polling is what an agent does when it has no way to be told. It burns calls on values that have not moved and still finds out late, because the interval was chosen by guessing rather than from how often the underlying thing actually changes. A price feed on a 180 second heartbeat polled every 10 seconds produces seventeen wasted calls and one that matters, and the agent cannot tell which was which.
What it would have to carry
The same verification block as a direct call: level, origin, source, cost, trace id. A subscription delivering unlabelled values would be a cheaper way to get exactly the problem Metera exists to remove, and no amount of convenience is worth that trade.
A refusal has to be deliverable too. If a scheduled collection comes round and nothing meets the level the agent asked for, the honest event is a refusal, not a silent gap and not the last good value sent again. A stale value that looks fresh is the exact failure this product exists to catch, and a scheduler is the easiest place in any system to introduce one by accident.
The hard part is deciding when to fire
Two obvious models, and neither is right alone. On a fixed interval, which reintroduces polling with extra steps. On any change, which turns a noisy source into a firehose and bills the caller for every flicker in the last decimal.
The version worth building fires on a change that matters: a deviation threshold the caller sets, plus a heartbeat so silence is distinguishable from a broken subscription. That is the same shape a real oracle feed uses, for the same reason, and it means the caller declares what a meaningful change is rather than us guessing on their behalf.
What it would cost
The way everything else is charged: on delivery, nothing on refusal. A subscription billed by the hour would start earning money for values nobody read, which is precisely the incentive we have spent the rest of the product avoiding.
What already exists underneath
The engine already runs scheduled collection internally, with a per-item spend ceiling and a minimum level per item. What is missing is the caller-facing half: a way to declare a subscription, a delivery channel, and a guarantee about what happens on the collections that fail. The on-demand call covers everything today, and for most agents a call at the moment of the decision is the correct pattern anyway.