Budgets and real money

Agents doing real work spend real money — model tokens on every run, and costs measured in currency where you track them that way. Budgets are how that spend stays a decision you made rather than a number you discover. This page covers where budgets attach, what happens as one fills up, and why hard caps matter more for a solo operator than for anyone else.

A budget attaches at the scope you care about

You set budgets in tokens or money, at whatever scope matches how you think about the spend: the whole workspace, a project, a goal, a single task, one agent, or one workflow. The scopes answer different questions. A workspace budget is the ceiling on the whole operation. A project or goal budget keeps one initiative from quietly eating the month. An agent budget bounds a specific worker while it is still proving itself, and a task or workflow budget caps one job at what that job is worth to you. Because the scopes stack, you can hold a broad workspace ceiling and still draw a tight line around the one experiment you are unsure about.

This is the money half of the control story: autonomy levels decide which actions need your sign-off, and budgets decide how much any of it may cost — and budgets hold at every autonomy level, including full autonomy.

Alerts arrive before the cap, and the cap actually stops work

A budget is not a report you check. It comes to you. At 80% of a budget you get an alert in your inbox — early enough to decide whether the pace is fine, the cap was too low, or something is burning tokens it should not. At 100%, work under that budget pauses. Not a warning while spending continues: the agent stops and cannot proceed under that scope until the budget changes.

A paused agent is not stranded, and you do not have to notice the pause yourself. The agent asks for more, and that request lands in your inbox like any other decision: what it was doing, what it has spent, and your call to make. Approve an increase and work resumes. Decline and it stays paused. Either way the spend decision was made by you, in the open, with the run's history a click away — never by an agent that kept going because nothing stopped it.

Hard caps are what make delegation safe for one person

For a solo operator, the cap is the feature. A company with a finance team notices runaway spend eventually. When you are the finance team, "eventually" is your own next glance at a dashboard — and an agent working nights and weekends can do a lot of spending between glances. A hard cap converts the worst case from "whatever accumulated before I looked" to "the number I chose, plus a request in my inbox". That bounded downside is what makes it reasonable to raise autonomy at all: you can let an agent run unattended precisely because unattended has a ceiling.

It also keeps costs legible as the operation grows. When each project, goal, and agent runs under its own budget, the question "what does this job cost us per month" has an answer you set and can read, not a total you reverse-engineer from an invoice.

Size the first budgets in two layers

You do not need a finance model to set the first numbers — you need two layers. The outer layer is a workspace budget sized at what you could shrug off in a month if every guess underneath it turned out wrong: the amount that would annoy you but not hurt you. It exists so that no combination of mistakes below it can compound past a number you already accepted. The inner layer is a tight cap on each thing you are unsure about — a new agent, a freshly installed workflow, one experiment — sized at what a single cycle of that job is worth to you. A weekly job that saves you an hour is worth an amount you can state. Cap it near that, not near the workspace ceiling.

Do not agonize over the numbers, because a wrong guess corrects itself. Too low, and the agent pauses and asks — a request in your inbox with the work and the spend attached. Too high, and the 80% alert still arrives before the cap ever matters. Budgets are a setting you tune, not a contract you have to get right the first time.

The usage view shows which caps to move

After a few weeks the guessing stops, because the usage view shows what each agent and each workflow actually costs per run. Read it against the caps you set. A budget that fills every cycle while the output keeps passing review is simply too small — raise it and stop paying the interruption tax. A budget that never climbs past a fraction of its cap is slack you can tighten around work you still want bounded. And a job whose cost per run drifts upward with no change in its output is the early sign that something in the workflow deserves a look — visible in the numbers before it is visible anywhere else. Right-sizing this way turns budgets from a safety guess into a statement of what each job is worth: set once, checked against reality, adjusted on evidence.

Budgets bound what agent work can cost. The harder judgment is what agent work is worth — when a stream of output has earned less oversight. That is a question of evidence, and Deciding when to trust agent output shows where the evidence lives.