Claude Duet: What Background Agents Mean for Agent Skills

Anthropic announced Claude Duet on October 21, 2026. It is a research preview of Claude as a proactive agent: it surfaces what needs your attention, takes on tasks in the background, and keeps working after you close the app.

That last point is the one that matters for agent skills. Duet does not start from a blank slate. Anthropic says it shares the same memories, skills, and connectors as regular Claude. In other words, the skills you have already taught Claude follow you into the background. Good skills get amplified. Sloppy ones do too — now while you are not watching.

What Duet actually is

Duet is a standalone research preview, available through a waitlist to Pro and Max subscribers in the US. It runs on Claude Opus 5.5, on its own cloud-based computer, with its own browser for sites that have no connector. You message it by text or voice, in one ongoing thread, from standalone iOS and Android apps or on the web.

Each morning it sends a short briefing drawn from your connected apps: what needs attention, what is waiting, what it has done. Through the day it checks those apps and messages you only when something requires action. Tasks range from the administrative — scheduling a meeting, checking for unpaid invoices, drafting a payment reminder — to the substantial: building a dashboard, making a forecast, prototyping, or building a web app.

Two guardrails are worth noting, because they define what "background" means here:

• It asks before it acts on your behalf. Anthropic says Duet is designed to check with you before sending a message, making a payment, signing a document, or deleting anything. Only you can mark a task complete.
• It is not free, even though there is no extra charge. Duet is included in Pro and Max plans, but background work counts toward your existing usage limits — and Anthropic warns that background activity can exhaust those limits sooner than chat alone.

Team and Enterprise plans are not eligible yet. Anthropic says it plans to expand access later. Treat every claim about Duet as preview-stage: availability, limits, and behaviour can change before general release.

Why this changes the value of a skill

In a normal chat, a bad instruction is visible. You see the output, you correct it, you move on. In a background agent, the same bad instruction runs while you are doing something else. The cost of a vague skill is no longer one bad answer — it is a morning briefing full of confident, slightly wrong work.

That flips the skill question. It is no longer "which skills sound impressive?" but "which skills can I trust unsupervised?" Three properties separate the two:

1. A skill should say what "done" looks like. "Write clean code" is not a skill; it is a wish. A trustworthy skill defines the outcome, the checks, and the failure modes: what to verify before handing work back, what to refuse to guess at, what to flag instead of fixing silently.

2. A skill should be tested, not just written. The strongest signal is a real task run through the skill with the result checked — a test suite that passes, a review checklist that actually caught something, a page that renders. If a skill has never been run against a real task, a background agent is the wrong place to find out.

3. A skill should be small enough to audit. The skills worth carrying into Duet are the ones you could read in a few minutes and explain to a colleague: what it does, what it will not do, and when it should stop and ask. A 2,000-skill bundle fails this test by construction. Nobody audits 2,000 skills, and a background agent will happily use the ones you never read.

That is also why the "skills follow you" detail matters more than the morning briefing. Your skill library is becoming portable behaviour — the same instructions, running in chat today and in the background tomorrow. Investing in a small set of verified skills now pays off in both places. Collecting a large set of unverified ones compounds the risk in both places.

What to do this week

If you use Claude Code or Claude regularly, Duet is a reason to audit your skill library, not to expand it:

1. List the skills you actually use. Most people rely on five to ten. Those are the ones that matter.
2. Read each one and ask: would I let this run while I am away? If the answer is no, either tighten it — add explicit checks and stop conditions — or stop using it.
3. Add one verification step to each keeper. A test command, a checklist, a "show me the diff before you commit" rule. Verification is what turns a skill from a prompt into a process.
4. Keep the rest out of the way. A skill that is installed but unused is still available to an agent. Remove what you do not trust.

None of this requires Duet access. The audit improves your chat results today, and it is the same work you will need if and when the waitlist reaches you.

The honest caveats

Duet is a research preview, not a finished product. It is US-only for now, waitlisted, and limited to Pro and Max. Its background checks consume the same usage allowance as chat. And the approval gates — asking before sending, paying, signing, or deleting — are exactly the places where a preview product will be tested hardest in the real world. Do not hand a background agent your inbox, your calendar, and your payment tools on day one and assume the gates will catch everything. Start with low-stakes tasks, read the morning briefing critically, and expand access as the product earns it.

The bigger story is not Duet itself. It is that Anthropic, OpenAI, and Meta are all converging on the same shape: a proactive agent that works in the background, on a virtual machine, from your existing skills and connectors. The agent layer is becoming a commodity. The durable asset is the behaviour you teach it — the skills, the checks, and the judgment about when to stop and ask.

Build that asset carefully. Duet just made it run while you sleep.

If you are starting that audit, begin with a small set of skills that are already documented, tested, and honest about their limits — our best Claude Code skills are picked on exactly that bar, and the full skills catalogue is there when you want to browse further.