What an AI Agent Actually Does All Day

"Agent" is the most overloaded word in software right now. Here is what one of ours — a Player, in our vocabulary — actually does during a shift, using a receivables chaser as the specimen.

06:00 — the watch

The Player queries the accounting system for invoices past their grace period. No inbox, no memory of yesterday's mood. A list of invoice records, aging by days. This morning: eleven.

06:02 — the judgment calls

For each invoice, it makes small decisions inside guardrails we wrote. First-time late or repeat pattern? Longtime customer or new account? Amount worth escalating over? It drafts a different note for the customer on invoice three of a long, previously punctual relationship than for the brand-new account thirty days silent. Tone rules are policy, not improvisation.

06:05 — the sends and the swerves

Eight notes go out. Three do not: one account is flagged as a sensitive relationship, one amount crosses the escalation threshold, one has a reply in the thread from yesterday — a human wrote in, so the Player stands down and routes the thread to a person with context attached. Knowing when *not* to act is most of what we build.

06:06 — the receipts

Every action lands in the run log: items processed, items routed to humans, tokens consumed, cost in cents. That log is what the client sees in their portal — "312 invoices chased, 47 paid after follow-up" — and what our monitoring watches. Three failures in a row and the Player is benched automatically, with a human paged. Blow past its monthly token budget and it pauses itself.

An agent is not a colleague. It is a process with judgment inside guardrails — and the guardrails are the product.

The part nobody demos

The demo is always the happy path. The production system is the swerves: the stand-downs, the escalations, the automatic bench. When you evaluate anyone's agent — ours included — skip the demo and ask to see the run log. The boring rows are where the trust lives.