Delivery
Delivery is the last axis and the one people underrate. It decides how a finding gets from a completed run to a person, and an agent whose findings arrive at the wrong moment or in the wrong volume is functionally an agent that isn't running.
It changes behavior indirectly but strongly, because it sets the cost of a false positive. An agent that sends immediately can tell you about a cancellation while the slot is still there — and can also wake you three times for nothing, after which you stop reading it. An agent that rolls up into one daily message is impossible to be annoyed by and useless for anything time-critical.
The rule that keeps it useful is that a run with nothing to say sends nothing. Silence has to be meaningful, or the signal is gone within a week.
Delivery in this deployment
What this deployment can send today.
| The agent's findings, only when there's something to say. | available | |
| Daily digest | One email rolling up every agent's day. | available |
| Slack | Findings into a channel via the Slack connector. | available |
How to choose
Ask what you would do in the first minute after the message arrives. If the answer is act now — a slot opened, a price dropped — send it immediately. If the answer is read it later, put it in the digest, where it costs you nothing to ignore.
Most agents belong in the digest, including ones that feel urgent when you build them. It is the setting you are least likely to turn off in a month.
One agent can reasonably be immediate; five agents all being immediate is a notification stream nobody reads. Delivery is worth deciding across your agents rather than one at a time.
There is no automated experiment runner yet. Nothing here sweeps this axis for you — changing it means editing the agent and reading the next run. A compare mode that runs one task across several compositions is named on the index as not built, and that is still true.