Autonomous agent: what happens to your budget mid-task?

You ask an autonomous agent to "compare thirty suppliers and write a summary". It works for ten minutes, or an hour. How much does that cost you? With this kind of tool, the question is not only the price of the plan: it is what happens when the task exceeds your expectations.

This article looks at a single point, budget handling mid-task, at Manus and at Ember. The two tools do not do the same job, and our overall comparison covers that choice. Here, we only compare how each one protects your balance.

What Manus’s help center says

Manus’s help page on credit consumption states that credits are spent across three items: language model tokens (planning, decisions, writing), virtual machines (a cloud environment for files, the browser and code) and third-party APIs. It adds that a task’s consumption is determined by its complexity and duration, and that credits are only consumed during active task processing.

On what happens when the balance runs out along the way, the page is clear: in rare cases, a task may complete an action whose cost slightly exceeds the remaining balance, which can show a temporarily negative balance, resolved at the next credit refresh. While the available balance is at or below zero, you cannot start a new task. Manus explains that it calculates and deducts each action’s cost at the start of the following action, which produces this overshoot.

The same text gives the order in which credits are consumed: event credits, daily credits, monthly credits, add-on credits, then free credits. It also specifies that purchases and refreshes follow UTC+0. The page read does not detail what happens when you stop a task yourself. It does state that Manus fully refunds the credits consumed by a task that fails for a technical reason on its side, and that completed tasks, and the storage or deployment of their outputs, consume no credits.

What Ember does

In Ember, a credit is an accounting unit for the real cost billed by the model provider. There is no price list per action: each call is charged at its real cost, right after it succeeds. Ember therefore does not announce a price before an action, and actual consumption can be read in the usage history.

For a Second Brain conversation, several rules resemble the Manus case, with differences:

• Once a turn is admitted, its image and Creation work runs to completion at measured cost, even if the balance becomes negative. The next turn needs credit.
• Stopping is explicit. Closing the browser does not cancel work that has already been admitted.
• A voluntary stop keeps the cost already incurred with the provider.
• A confirmed failure with no saved answer cancels the debit. While an operation’s outcome is uncertain, Ember does not announce a refund.
• In a team, the balance is shared and the owner or administrator sees each expense with the member’s name.

Four questions before delegating an open-ended task

1. What sets the end of the task? Write down the expected deliverable (a ten-row table, not "a study"). A vague brief lengthens the duration, and therefore the cost.
2. Where can I stop it, and what is lost? Find the stop button before launching, not during.
3. What is kept if it fails? Check whether partial results are kept and whether the failure is billed.
4. Who pays in the team? A shared balance with no per-person limit exposes the team to one person’s task.

In practice, a founder mainly trying to settle a decision (hire now or in six months, accept a client) has little reason to let a web agent run for an hour. Conversely, an open-ended collection of data from the web belongs to an execution tool, with a cap you set yourself.