An Agent Quietly Burned Through My Budget Overnight

An agent stuck in a loop does not announce itself. It retries, spawns follow-up work, and burns tokens overnight, and the first signal you get is an invoice. For a solo operator funding everything personally, that is enough reason to turn agents off entirely, so the runaway loop costs you twice: once in money, and again in everything the agents would have done for you afterward.

A usage graph is not a limit

Most tools answer this with a usage graph, which is informative at exactly the moment it is too late. Bounded spend needs a hard cap the system enforces, and a chart you promise to check is no substitute.

Task Machine budgets are hard limits. You can set them in money or tokens for the whole workspace, a project, a goal, a task, a single agent, or one workflow, and you get an inbox alert at 80% and again at 100%. When a budget runs out, new agent work under it pauses instead of carrying on with a notification.

Agents that need more have to ask

What makes caps livable is what happens at the limit. An agent that needs more budget asks for it. The request lands in your inbox like any other approval, with what you need to decide: what the agent is working on, what it has spent, and how much more it wants. You grant the increase in a click, or you don't.

That turns the nightmare scenario into an ordinary Tuesday. The loop that would have run all night hits its cap at 11 p.m., pauses, and asks, and you wake up to a request instead of a bill.

Budgets make delegation cheap

Hard caps let you delegate more. When the worst case on any experiment is bounded, say this agent can spend at most twenty dollars before anything needs you, trying things stops being a financial decision. You hand over more work because the downside is fixed and visible, and the spending that does happen shows up in your usage view, attributed agent by agent and workflow by workflow.