Summary
Anthropic has opened a US waitlist for Claude Duet, a research-preview agent for Pro and Max subscribers. Users can hand it tasks by message; Anthropic says it can continue on a cloud computer and return when finished or when it needs a decision.
The operational question is whether that work can run usefully without repeated intervention. Duet checks connected accounts every few hours, and background activity draws on the subscriber’s existing Claude usage limits. Anthropic has not published independent task-completion or reliability results, so the preview’s advertised capabilities do not yet establish how much work it can finish in practice.
What happened
Anthropic describes Duet as a proactive agent powered by Opus 5.5, available through a US waitlist for Pro and Max subscribers. It can carry over Claude memory, skills, and connected apps, and run tasks on a cloud computer with its own browser while the user’s device is offline. The company lists examples such as preparing briefs, drafting replies, creating documents, and following up on invoices.
The product’s operating model has two notable constraints. Anthropic says Duet checks connected accounts every few hours rather than continuously, and that background activity shares the user’s plan limits, with usage counted differently from a typical conversation. It also requires approval for sensitive actions such as sending messages or making payments. These details describe the intended product; the source provides no independent evidence of how often tasks finish, how much intervention they require, or how plan limits affect completion.
What the archive adds
The archive adds limited but relevant context. According to dossier finding d2, Anthropic’s research on multiagent systems examines conflicts and failures when agents operate in shared systems. That does not establish how Duet performs, but it reinforces that delegation involves coordination and control as well as model output. The archive’s coverage of Anthropic’s negotiation research likewise describes questions about when an agent should seek human feedback; it is exploratory work, not evidence that Duet can negotiate or complete consequential tasks reliably.
Trend signal
Duet represents a move from answering a prompt to continuing a task after the user steps away. Yet persistence alone is not a measure of autonomy: an agent that checks in periodically, hits usage limits, or often needs decisions may still require frequent supervision. The practical distinction is between work that can be queued and completed asynchronously, and work that remains a sequence of interruptions. The preview’s design raises that distinction, but its public materials do not resolve it.
Why now
The primary evidence says Duet can continue on a cloud computer while the user is away, which makes persistent delegation a product focus. Its described tasks span multiple connected services, while sensitive actions require approval; that combination raises a practical need to know how much work completes before a person must intervene. According to dossier finding d2, Anthropic has also studied conflicts and failures in multiagent systems. These facts help explain why delegation boundaries matter now, but they do not establish Duet’s performance or Anthropic’s motive.
Connections to other events
The hype check is straightforward: a cloud computer and broad app access show that Duet is designed to work beyond a single chat turn; they do not show that it can reliably finish a workflow unattended. Periodic account checks may make the agent less immediate than a continuously watching assistant, while shared usage limits may cap the amount of background work available. Our read is that the decisive measure will be completed tasks per user interruption, not the number of integrations. That is an inference from the operating constraints, not a result Anthropic has reported.
Why it matters
Builders evaluating delegated agents should test complete workflows, not just whether an agent can access an app or produce a plausible draft. Record how often it finishes, pauses for a decision, or needs correction, and how much plan capacity the run consumes. Buyers should treat advertised tasks as candidate use cases rather than proven capabilities, and start with work where a delay or an unfinished run is recoverable. The important product question is whether the agent reduces supervision time after accounting for its checks and requests for help.
What happens next
- Anthropic will publish additional guidance about how background work uses Duet subscribers’ plan limits. (6 months, 55%). Watch for: Updated Duet or plan documentation explains background-task usage in more specific terms.
- Anthropic will report at least one concrete Duet performance measure, such as task completion or user intervention, as the preview develops. (12 months, 40%). Watch for: A product update or public report provides a task-success, intervention, or reliability figure.
Limits and uncertainty
This is a research preview, and the evidence does not give task-success rates, intervention frequency, independent reliability or security measurements, or a separate price. It also does not quantify how the shared usage limits affect real workloads. Anthropic’s stated checking cadence and product examples describe its offering, not independently verified performance. This take could be wrong if typical tasks complete with little intervention and modest usage, or if the account-check schedule rarely affects the workflows users care about.
Skill to build now
Practice measuring delegation overhead. First exercise: give an agent one low-risk task that can be completed asynchronously, then record completion, pauses for user input, corrections, elapsed time, and usage consumed. Compare the result with completing the task manually.
Reader takeaway
This week, run a small, measurable delegation test on a low-risk task that can wait, such as preparing—but not sending—an invoice follow-up. For each run, log whether it finished, how many times it requested your input, how much correction was needed, and the usage consumed. Compare the total time spent supervising it with the time the task normally takes. That will show whether the agent is actually removing work or merely moving it into a review queue.
Open question
What level of user intervention and shared plan usage will people tolerate before an asynchronous agent feels like another inbox to manage?