Limits and costs

An agent decides its own route, so the interesting question is not what it will do but what it cannot exceed while doing it. Everything below is a bound, not a forecast.

The two numbers you agree to

Per run. The most one run can cost. When a run reaches it, the run stops and hands back what it has, rather than continuing.

Per month. The most the agent can cost you across a calendar month. This is the enforced one. When it is reached the agent refuses to start and tells you so:

"Competitor watch has used its $22.50 monthly budget. It will run again next month, or you can raise the ceiling."

Both are shown before you create anything, because either one alone misleads. The per-run figure hides that a daily agent runs thirty-one times. The monthly figure hides what a single run can cost.

The monthly figure is derived from the per-run ceiling and the cadence. Change the schedule and it moves. An agent with no schedule is given an allowance of ten runs a month rather than an unlimited one.

What else is bounded

Bound	What it stops
Tool calls per run	An agent looping on a page that never quite answers the question
Tool calls per tool	Twenty searches and one page read, when the point was reading pages
Concurrent runs	A schedule firing while the previous run is still going
Which tools it may use	An agent reaching for something its shape never needed
Which sites it may reach	Only relevant where an agent also reads your own data. See What agents can see

A failed call still counts against the ceiling. An agent that spends ten attempts on an unreachable page has used ten attempts, and the alternative is an agent that retries forever.

Running one twice at once

If you start an agent that is already running, you join the run in progress. It does not start a second one, and nothing extra is spent. Studio says so explicitly:

"Already running. Joined that run rather than starting a second one."

This holds across every surface. Pressing the button in Studio, asking an assistant over MCP, and the schedule firing all reach the same check.

When an agent asks for help

Some shapes split their work: several parallel investigations, then one pass that reconciles what they found. Two rules govern that split.

The reconciling pass is paid for first, before any of the parallel work is funded. The obvious approach, handing each investigation whatever is left, fails twice: the first one can eat everything, and nothing remains to make sense of the results with.

If the remaining budget will not fund a useful share each, the agent does fewer investigations rather than starving all of them. Three that finish beat seven that all stop at forty percent.

When it could not do everything

An agent that had to cut back says so, on the agent and in the run:

Did less than asked. The last run could not do everything it was asked. The next one will try a little less rather than repeat the same over-reach.

This is surfaced rather than logged quietly, because an agent that silently researched three of five sources looks exactly like one that found nothing worth reporting. Those are very different results and you should not have to guess which you got.

The next run then starts smaller, on its own. Repeatedly degrading usually means the subject is too broad for the ceiling: narrow the subject, or raise the ceiling.

Your organisation's limits still apply

An agent's own ceilings sit inside your organisation's overall AI budget and quota. Before an agent run starts, Studio checks that there is enough budget left for the whole run, not merely that some budget remains.

That distinction matters. Starting a five dollar run against ten cents of remaining budget means the run begins, spends the ten cents, is cut off part-way, and produces nothing usable. You would be charged for work you cannot use, and it would look like a fault rather than a budget. So a run that cannot finish is not started.

A scheduled run refused this way is not retried and does not count against the agent's failure allowance. You are notified once, not once per firing.

Changing a ceiling

On the agent in Studio, or by asking an assistant connected over MCP. Raising a monthly ceiling takes effect immediately; the agent will run again on its next scheduled firing, or when you run it once by hand.

Next

• What agents can see: the read side of the envelope
• Testing an agent: reading what a run actually cost and did