Oroboro LabsThe attention tax: what nobody tells you about 24/7 AI agents
An agent that runs around the clock is sold as passive income. It isn't. It's a business that pays you nothing and bills you constantly — in the one currency you can't mint more of. We run such agents daily; here are the three bills nobody itemizes, and what legally reduces them.
Bill 1 — the context toll
A tool that stays connected to your agent publishes its own definition in every request, whether or not that task uses it. Five servers left on is a toll collected five times per call. The fix is embarrassingly mechanical: prefer the tool that doesn't live in the context, and turn off what the task doesn't touch. We measured the same effect on a bigger scale with documents — reading a derived index instead of a raw vault made agent reading ~150× cheaper (the measurement is here). Same principle, different layer: what you keep loaded, you pay for on every turn.
Bill 2 — interruptions that train you to ignore
An agent that notifies you of everything is training you to notice nothing. A notification that doesn't demand a decision is noise wearing a productivity costume. This one is qualitative — we have no counter for attention residue, and we won't invent a number (rule in our house: a number without a method doesn't travel). But the design fix is structural: batch reporting. One morning digest of what failed and what it cost beats a ping per event, for the same reason cron beats a daemon spinning in your peripheral vision.
Bill 3 — the review step that doesn't shrink
The worst one, because it can't be automated away honestly. A 24/7 agent produces 24 hours of output; reviewing it honestly also takes time, and if you don't budget for that, you haven't automated the work — you've relocated it to 2 a.m., where nobody audits it. Our answer is to make review sampled and adversarial instead of total and sleepy: a second agent whose explicit job is to reject the first one's work, with checks that produce a count (coverage: what's missing and why; fidelity: sample the claims against the source). Auditing a rejection attempt is much cheaper than vigilance.
Paying less, without lying to yourself
- Context: a tool not in use for this task stays off. Between two equivalent options, prefer the one that doesn't sit in the context — a CLI called on demand costs nothing while idle; a connected server bills per request.
- Attention: batch reports on a schedule, not pings per event. One item per line, each with its evidence path.
- Review: give the agent a written contract — it logs its own failures with reasons and keeps going. You audit the record once, not the day.
The punchline: automating what you can't afford to audit isn't scale — it's debt, with interest collected in sleep.
If you run an always-on agent: which of the three bills eats you the most — and did you ever catch bill 3 in the act?