Claude Code Can Now Run for Hours Without You

You have not been using AI wrong. You have been using it small.

What is Claude Code's /goal command?

/goal is a command inside Claude Code, Anthropic's coding agent, that gives the AI a finish line instead of a to-do list. You type the condition once, for example "do not stop until every unit test passes," and Claude keeps working, turn after turn, until a second, separate model checks the transcript and confirms the condition is actually true. It shipped in Claude Code version 2.1.139 and it changes one specific thing: you stop being the loop. The AI is.

Thariq Shihipar, who works on the Claude Code team at Anthropic, put it plainly in a July 2026 interview with creator Peter Yang: "As the models have gotten smarter, they need less direction, fewer constraints, and fewer examples. We've got /loop, we've got /goal, and we've got workflows. These are all geared at trying to get the agent to run for long periods of time."

Three commands, three different jobs. /goal starts a new turn the moment the last one finishes, and only stops once an evaluator model says the condition is met. /loop starts a new turn on a timer, useful for something you want checked every hour rather than finished once. Workflows go further still: Claude writes a small script that spins up several sub-agents at once, each doing a slice of the work, with a separate agent grading the results before anything ships.

The demo that makes it click, and it is not a coding demo

Here is why a business owner who has never opened a terminal should care. On the call, Thariq recorded a rough video of himself talking to camera, ten minutes before the interview started. Then he gave Claude Code one prompt: transcribe the recording with Whisper, use Remotion to build a UI that highlights each word as it is spoken, add an overlay where he pointed, fade to black at the end, and "/goal, don't stop until the video is fully rendered."

One prompt. No back and forth. Claude transcribed the audio, built the caption overlay, rendered the video, and did not hand control back until the file existed and played. That is the whole shift. Not smarter answers. Fewer handoffs.

Swap "render a video" for whatever eats your afternoons; the mechanism is the same.

Why this matters if you have never written a line of code

Picture Amara. She runs a fourteen-person skincare brand in Lagos, shipping across Nigeria and into three neighbouring countries. Every week she has a growing pile of unanswered reviews on Google, Instagram, and her Shopify store. Some are five stars and easy. Some are a customer whose order arrived late, and those take real thought to answer well. She has been doing this in stolen ten-minute blocks between everything else, which means the ones that need care the most get answered last, if at all.

A goal-driven agent does not need Amara to open each review and write a reply. She gives it the finish line: "Draft a reply to every review from the last 30 days that does not have one yet. Match our brand voice, acknowledge the specific complaint before offering a fix, flag anything that mentions a damaged or missing item for my review before sending. Don't stop until the backlog is at zero." Claude works through the list, turn after turn, checking itself against that condition until an evaluator model confirms the backlog is actually empty, not just "probably fine."

That is the part worth sitting with. The AI is not trusted to grade its own work. A second model checks it.

What workflows add that /goal cannot do alone

/goal is built for tasks with a clear, checkable finish line: tests pass, backlog is empty, file exists. Google's own AI Overview on this feature is explicit that you should avoid it for "subjective work that lacks a measurable done state, like improve the onboarding experience." Fair. Not every business task has a pass or fail button.

That is what workflows are for. A workflow is Claude writing a short script that fans a task out across several sub-agents running in parallel, each one working a piece of it, with a separate verifier agent checking the output against a rubric before it counts as finished.

Say Amara needs product descriptions written for 40 new SKUs before a launch. A workflow assigns each description to its own sub-agent, then routes every draft through a second agent holding the brand-voice rubric: right tone, no banned words, correct claims about ingredients, length under 150 words. Nothing ships until it passes that second check.

Thariq's team has a name for the failure mode this avoids: "self-referential bias." A model grading its own homework tends to be generous with itself. Split the writer and the checker into two separate agent runs, and the check actually means something.

Isn't this just a chatbot that runs longer?

No, and the difference is the evaluator. A chatbot that you leave alone for an hour will happily decide it is done long before it actually is, because nothing is checking its claim. /goal and workflows both add a second, independent pass: a separate model reading the transcript, or a separate model grading the output, before either one is allowed to call the job finished.

That second check is the whole point. It is also why /goal works best when you can state the finish line in a way Claude's own output can prove, the way "all tests pass" is provable from a test run, and "the tone feels more premium" is not. Write your condition so it can be demonstrated, not just asserted.

How to try this in your business this week

1. Pick one backlog with a real finish line. Unanswered reviews, unread support emails from today, unpaid invoices past 30 days. Something you could, in theory, get to zero.
2. Write the condition like a contract, not a wish. Not "handle the reviews," but "reply to every review from the last 30 days that has no reply, matching this tone, flagging anything with a damaged-item complaint before sending."
3. Give it the goal and let it run. In Claude Code, that is literally typing /goal followed by the condition. Check back in an hour, not every five minutes.
4. For fuzzier, many-item tasks, ask for a workflow instead. Say what "good" looks like as a short checklist, and ask Claude to spin up a workflow that does the work and checks it against that checklist before it counts as done.
5. Read what it produces before you trust it wholesale. Thariq's own warning: "the prompt box can definitely just be a lazy button." Skim the first batch closely. That is where you catch a badly worded condition before it repeats the mistake forty times.

What you have after this

A week in, you have one backlog that used to sit ignored, now cleared on a schedule instead of when you finally have an hour. A month in, you have written two or three conditions well enough that you barely think about them, they just run. Six months in, the list of things you personally have to remember to do gets shorter, not because you got faster, but because you stopped being the one holding the loop open.

Same business. Same team. A meaningfully lighter Tuesday.

How Notma helps you find the first one

Most owners we work with have three or four backlogs like Amara's sitting in plain sight, invoices, reviews, onboarding emails, none of them urgent enough to fix today and all of them heavy enough to notice. The 14-Day AI Readiness Sprint exists to find the one with the clearest finish line and get an agent running against it inside two weeks, not a quarter.