How to Pace Your AI Agents' Usage So You Don't Get Locked Out (AICEP v0.2 draft)

AICEP v0.2 draft. Provided "as is", without warranty. Normative rule: Usage Pacing (§10.6). Free prompt: Allowance Pacing Plan (runbook G12).

Most AI assistant platforms give you a usage allowance that resets weekly. Run one assistant and you'll rarely notice it. Run a team of agents, with builds, digests, monitors and reviews all drawing from the same pool, and you can burn a week's allowance by Wednesday. Then everything stops, including the parts you actually needed.

Pacing is how you avoid that. It isn't about using less. It's about spending evenly, protecting a floor, and keeping a reserve for the day something breaks.

The core idea: a daily share of the week

If the allowance resets every seven days, each day gets about one seventh, roughly 14%. Your pace on any day is simply days since reset divided by seven. Example: three days after reset, pace is about 43%. If your meter reads 40%, you're on track. If it reads 60%, you're ahead of pace and need to slow down now, not on day six.

That's the whole measurement. You don't need a spreadsheet. Nearly every platform shows a percent-used meter, and checking it once a day at the same time is enough.

Set four limits before you build

Daily ceiling. Your normal day, around one seventh of the week. Above it, only urgent work continues.

Hard stop. About 1.5 to 2 times the ceiling. Past it, everything stops except keeping the lights on.

Reserve. About a third of the weekly allowance that planned work never touches. It's for the outage, the urgent fix, the surprise.

Floor. The minimum that must keep working every day: you can reach your coordinating agent, urgent items get handled, core scheduled jobs run.

(All figures here are illustrative examples, not measurements from any real deployment.)

When you're ahead of pace: pause optional work first

Sort work into floor, urgent and optional. When the meter runs ahead, optional work pauses first: new builds, extra digests, polish, repeat checks. Give one agent, ideally your security agent, the job of checking the total daily and calling the pause for the whole team, then lifting it once you're back on pace. One shared budget, not a budget per agent, because agents don't see each other's spend.

Build heavy, release light

Let big builds run heavy inside a bounded window, one or two sessions per feature. If it spills over, stop and re-plan. Finished work goes into a queue and ships once a day in a fixed slot, in small batches, checked by a script (health call, count, a sample) rather than an agent clicking through pages taking screenshots.

The refill window

Allowance you don't use before reset is usually gone. So in the last day before reset, if you're at or under pace, let spare capacity drain the backlog, most valuable items first. Leave the reserve alone until the final hours, and stop at reset or the hard stop, whichever comes first. If you aren't sure exactly when reset happens, stop earlier.

Cheap habits and expensive ones

Route by model tier. Routine runs, status checks and scripted jobs go to the cheapest model that does them reliably. Save the top tier for hard reasoning. When a fixed procedure fails, escalate to a human, not to a bigger model.

Cheap: short results from background workers; long output written to a file and linked; fresh sessions at milestones with a short handoff note; one dispatch per workstream; checks at the lowest useful frequency.

Expensive: very long sessions that re-read big files; pasting large output into chat; re-verifying things that already passed; browser crawls and screenshots for routine checks; hourly monitors that could be daily.

Optional: the usage export

Many platforms let you download a usage export as a CSV. It's useful once a week to see which jobs cost the most and where to trim. But it's a tune-up, not the steering wheel. Daily pacing works fine from the meter alone, and a plan that only works when someone remembers to download a file will fail on the busiest week. Keep raw export data out of chats and public pages.

What it buys you

Without pacing, the week ends whenever the allowance does. With it, Tuesday's slowdown of optional work replaces Thursday's lockout, and the work that matters keeps running all seven days.

The rules are in Usage Pacing (AICEP v0.2 draft §10.6), and the ready-to-use prompt is in the runbook: Allowance Pacing Plan (G12).