Claude Duet: Work Continues After You Close the App
Claude Duet explained: cloud execution, Skills, connected apps, preview access and usage limits, with a practical example of a recurring research task.
On this page
Based on official materials checked October 22, 2026. This article is an analysis, not a hands-on Duet review.
Close the app. The work continues. That is the promise behind Claude Duet, which Anthropic announced on October 21, 2026.
For people who already build Skills, the interesting question is what happens to a reusable workflow when it can run while its owner is away. A good answer needs more than a list of available tools: it needs a clear assignment and a result someone can check.
What is available at launch?
The product page describes the following starting point. Treat these as the research preview terms checked on October 22, rather than a promise about future availability.
| Item | Research preview |
|---|---|
| Model | Claude Opus 5.5 |
| Access | Waitlist for Pro and Max subscribers in the US |
| Usage | Included during the preview; shares existing plan limits, including background checks |
Sources: Claude Duet: product and FAQ
Why closing your laptop does not stop the task
Duet runs on a cloud virtual machine with its own browser. Your laptop or phone can be off. It checks connected accounts every few hours and asks before sensitive actions such as sending messages, paying, signing or deleting.
The place where execution happens matters. A tab is an interface; a running process does the work. Closing the interface and shutting down the machine that runs the process are different events. Duet puts that execution environment in the cloud.
Periodic checks also change how to write an assignment. A daily digest can tolerate a delay. A task that depends on an immediate response needs a suitable trigger and a clear expectation about timing. An agent described as always available should not be assumed to watch every change in real time.
Sources: Introducing Claude Duet
Where Skills fit into a background agent
Duet can use the memory, Skills and connected apps in your Claude account. Its product page lists more than 6,000 connectors and plugins combined; that figure is not a count of Skills.
Anthropic describes a Skill as a package of instructions, metadata and optional resources, such as scripts or templates. It can give a recurring task a consistent method and output format.
Consider a weekly reading brief. The connector provides access to the material. A Skill can specify how to group it, preserve source links and flag uncertainty. The execution environment runs the steps. Memory may supply context, but the assignment should still say which material belongs in this particular brief.
Our interpretation is that background agents make reusable task definitions more valuable. A workflow that only says ‘write a good summary’ leaves too much to guess. A workflow that defines its inputs, evidence and completion conditions is easier to delegate and review. This is a design implication, not a claim that Duet is the first official use of Skills.
Sources: Claude Duet: product and FAQ · Agent Skills overview
How Duet differs from Claude Code Mods
Announced October 1, Claude Code Mods are small TypeScript functions that customize Claude Code behavior and interfaces. Duet is a hosted agent to which users delegate work. The two address different parts of a workflow.
When comparing tools, start with the job. Customizing a coding environment is one job. Handing over an ongoing task is another. A feature that helps with the first does not by itself establish where the second runs, how it resumes or how you inspect its result.
Sources: Claude Code Mods
An example: a weekly research brief
Here is a task brief you can adapt. It is our proposed example, not a tested Duet recipe or a description of a specific product screen. Connect and authorize the required sources before using it in an agent.
Task: Prepare a weekly research brief.
Inputs: Only the material in the folder I have authorized.
Period: Items added since the last completed brief.
Output: A draft grouped by topic, with a source link for every factual item.
Review: Separate source facts, your interpretation and missing information.
Completion: Save the draft and report which items were included or skipped.
Escalation: Ask if access is missing or the requested period is unclear.
Delivery: Keep the result as a draft; ask before sending or publishing it.Make the result easy to check
The example deliberately asks for a record of included and skipped items. A polished summary can hide an incomplete search. That small record lets the reviewer check whether an important document was absent, inaccessible or simply judged irrelevant.
It also separates evidence from interpretation. ‘The release notes list this feature’ and ‘this may change how our team works’ belong in different parts of the brief. Both can be useful; only the first is a fact established by the source.
For the first run, choose a bounded set of documents and compare the draft with that set. Check one source link, one excluded item and one conclusion. Expand the assignment after the output is useful. Installing more Skills before doing this check does not establish that the task works.
What this means for DSH workflows
DSH users can apply the same task-writing habits without assuming the products share an execution model. For any local or remote agent, identify the host that runs the process, the data it can read, the output location and the condition that stops the task.
A reusable Skill should help someone answer a practical question: if I leave this task alone, what will be ready when I return? Name the artifact, preserve its sources and say when the agent should ask for a decision. Those details make delegation easier to assess than the number of installed tools.
Start with one existing task, such as a release-note brief. Define its input folder and acceptance criteria, run it once, then review the saved result. That gives you a useful basis for deciding whether it should become a recurring background task.
Claude Duet questions
Can I close my laptop?
Duet’s execution runs in the cloud, so your device need not stay on. Official source
Does the 6,000 figure mean 6,000 Skills?
No. The product page counts connectors and plugins together. Official source
Should every Skill become a recurring task?
Start with a task that repeats, has accessible inputs and produces a result you can review. A one-off investigation with an unclear scope needs further definition before you schedule it.
What should I check before handing work over?
Check where it runs, what it can read, what it must produce and when it should ask you. Test that assignment on a small input set first.