Paying for silence
To put a bot in a conversation, the simplest design wakes it on a timer: every minute, look, maybe reply. It works. It also pays most where there is least to do.
The timer is always running
A session that wakes every sixty seconds is never off. Each wake reloads what it needs to know and spends its budget whether or not anything has happened. In a day that is well over a thousand wakes. A conversation that lasted ten minutes was watched for twenty-four hours. Almost every look finds an empty room and costs the same as a look that finds a message.
Conversation is bursty; polling is even
Talk comes in bursts with long silences between them. A fixed timer spends evenly against a load that is anything but. You pay the most during the quiet, which is most of the time. That is the wrong way round: the bill should follow the conversation, not the clock.
You cannot tune your way out
Poll faster and the bot feels present, but the waste grows with the rate. Poll slower and you save, but a reply now lands a minute late, which in a live exchange feels broken. A single interval forces a choice between cost and presence, and neither setting is right for both the busy minute and the idle hour.
Let the message do the waking
The better signal is the message itself. When something arrives in a room the bot owns, that arrival wakes it. Now cost tracks activity: you pay per message, not per minute, and silence is free. This is push, and it is how a messenger is supposed to work — nobody refreshes their inbox every sixty seconds.
Awake when it matters, asleep when it doesn't
Push is the spine, with two refinements. While a conversation is live, the bot stays close and answers in near real time. When it goes quiet, it stands down — no heartbeat, or a slow one — until the next message calls it back. A rare, cheap poll sits underneath as a safety net, for the case where a push is missed.
The principle is simple: presence should cost what attention costs, not what waiting costs. A design worth keeping makes waiting free.