Build · founder · 7 min read

Your Agent Can Now Run for 10 Hours Without Asking You Anything

Lovable's /goal runs up to 10 hours without pausing for questions, and costs noticeably more. Cursor shipped the same thing 11 days earlier.

Published September 3, 2026 ·

lovable cursor agents credits pricing founders

On August 30, Lovable shipped /goal. You type it at the start of a build message, describe what you want, and the agent works on that one message for up to 10 hours, without pausing to ask you questions, until it decides the goal is achieved.

Eleven days earlier, on August 19, Cursor shipped /goal — “a long-lived objective to work towards until it’s fully complete.”

Same word, same month, two vendors who are not usually copying each other’s release notes. When that happens it is worth stopping to ask what changed underneath, because it is not the feature. It is the unit you are billed in.

The unit of work just got 200x bigger

Everything about how you budget for these tools was built around a message. You send a prompt, the agent does a thing, you look at it, you send another prompt. Costly, but legible — you see the result before you spend again. Every pricing complaint about Lovable and Cursor over the last year has been a complaint about how much a message costs, and every defence has been “but you saw what you got.”

A goal run breaks that. Lovable’s own changelog is unusually plain about it: a goal “costs noticeably more than a regular build message.” Ten hours is not a longer message. It is a working day of an agent making its own decisions about what to try next, and each of those decisions is a billable step you did not authorise individually.

This is the seventh billing lever this site has tracked. The others — allowance size, metering unit, overage default, routing incentive, retention window, training rights — all changed the price of a thing you asked for. Goal runs change who decides how many things get asked for. That’s a different category of exposure, and it is the first one where the answer to “why did this cost so much” is legitimately “because it was working.”

The brake exists, and you have to set it yourself

Here is the part worth your five minutes.

Lovable’s credit check-ins — shipped August 17, before goal runs existed — are the only thing standing between a /goal and your balance. Lovable says so directly: credit check-ins “still pause the work at your check-in level” during a goal run. A single message can run for up to 10 hours, so a long run may check in several times along the way.

The check-in level is a whole number between 20 and 100,000 credits, set once in account settings and applied across every workspace you’re in. The default is 20.

Three things follow from that:

• If you ever dismissed a check-in with “Don’t ask again,” you have no brake on goal runs. Go turn it back on before you use /goal for the first time.
• The level is per-account, not per-project. The threshold you set because you were being careful on a client build also applies to the throwaway prototype, and vice versa.
• A check-in is a pause, not a cap. A run can drift slightly past the level before it stops. On a 20-credit check-in that’s noise. If you raised the level to something comfortable for a big build, the drift scales with it.

Cursor’s version has a different shape and the same underlying issue: goal runs pair with subscriptions (agents that wake on a PR or a Slack thread) and with Auto’s per-model metering, live since August 24. An agent that decides its own request volume, metered at a rate that varies per step, working toward an objective you set once. That is three sources of variance stacked, and none of them are the seat price.

What a goal run is actually good for

None of this means don’t use it. Used deliberately, /goal is the best thing either tool shipped in August, because it removes the failure mode that wastes the most founder time: the agent stopping to ask you a question you don’t know the answer to.

You know the moment. You’re four prompts into a build, the agent asks whether you want optimistic updates on the mutation, and you have no idea, so you say “whatever’s standard” and it burns two more turns interpreting that. A goal run just decides. For a founder without a technical co-founder, an agent that makes reasonable calls and shows you the finished thing is often more useful than one that asks perfect questions.

The tasks it fits:

• Finishing something you’ve already scoped. “Build out every CRUD screen for the admin section using the patterns in the customers page.” The pattern exists, the judgment is already made, you just want the volume.
• Grinding through a category of defect. Broken states, missing loading spinners, unhandled errors on every form. Boring, mechanical, and you’d never get to it manually.
• Anything you’d otherwise have paid a contractor a day for. That’s the honest comparison. A goal run that costs $40 in credits and replaces a $400 day is a good trade even when it goes sideways.

The tasks it does not fit:

• Anything where you’d change your mind halfway. The whole value proposition is that it does not check in. If you’re still deciding what you want, you are paying an agent to build the wrong thing for several hours.
• Anything touching money, auth, or customer data on a live app. Not because the agent is bad at it, but because “it ran unsupervised for six hours” is the worst possible sentence in a post-incident write-up.
• Your first hour with the tool. Learn what a normal message costs before you commission a working day of them.

How to run your first one safely

1. Check your check-in level is on and set somewhere you’d notice — Settings, then the credit check-ins control. If you turned it off, turn it on.
2. Screenshot your credit balance before you start. You want a before number that isn’t a memory.
3. Scope the goal narrowly and write it long. Counter-intuitive, but a goal run rewards specificity in a way a chat message doesn’t — there’s no follow-up turn to correct course. Name the files, name the pattern to copy, name what “done” looks like.
4. Say what it must not touch. Auth, payments, the database schema, whatever. It won’t ask.
5. Run it when you’re around. Not overnight. The first time, at least, you want to be able to stop it.
6. Check the spend against the result before you run a second one. One data point is enough to know whether this is a $15 tool or a $150 tool for the kind of work you do.

The thing to watch

Both vendors shipped a long-horizon primitive within two weeks of each other, and both attached it to a metered balance rather than a flat rate. That is not a coincidence — it is the only way the economics work, because a 10-hour run genuinely costs the vendor 10 hours of inference.

What it means for you is that the gap between your cheapest month and your most expensive month is about to get much wider, and the variable is no longer how much you used the tool. It’s how many times you said “go” and walked away.

Set the check-in. Then go build something.

Sources: Lovable changelog, Lovable goal runs documentation, Lovable credits and usage, Cursor changelog, August 19.