Andrew Warner and Andrew Kordampalos ask the right question about Claude Duet: does a polished, proactive assistant actually replace the other agents people already use? Their episode brings together early-user clips from Claire Vo, Peter Yang, Brock Mesarich and Riley Brown, then tests the product's organizing ideas against real work. The strongest evidence is for capturing tasks, resurfacing useful context and turning data into editable artifacts. The weakest point is handing off a browser task when a human must intervene.
Duet is promising for an individual who already works in Claude: one conversation, topic channels, visible tasks, scheduled check-ins and useful documents or dashboards. It is not a team of separate agents. It does have a cloud computer, but you cannot view or take over its browser, and it cannot operate your local desktop. Access is a US-only Pro/Max waitlist preview, with background work drawing from the same Claude usage allowance.
Watch Andrew and Andrew's Review
Viewing note: This is a commentary and demo-reaction episode, sponsored by Zapier. Some clips belong to other early-access reviewers. Their specific results are examples, not a controlled comparison among Duet, Grok Bot and Muse.
One Agent, Three Places to Put Work
The hosts like Duet's channels because they let a user separate sales, content or other topics without starting from scratch. Anthropic's help center says channels share the same underlying memory and Claude account context. That makes them topic folders for one assistant, not separately permissioned employees with different inboxes and identities. Andrew would rather give prospecting, client delivery and inbox triage to distinct agents, each with its own skills and boundaries. That is a preference about operating model, not evidence that Duet cannot do any of those jobs.
| Surface | Use it for | Where confusion begins |
|---|---|---|
| Main thread | Quick requests and the ongoing relationship with Duet. | Important work can disappear in a long conversation. |
| Channel | Keeping one area, such as email marketing, together. | It still shares Duet's memory; a channel is not a separate agent or data silo. |
| Task thread | Following a specific job, its progress and requests for review. | A task can sit within a channel, so users must remember where its discussion lives. |
That last point is the hosts' information-architecture criticism. A main chat, channel chat and task chat can each look like “the conversation.” Start with a few durable channels, then rely on the Tasks tab for work with a clear finish line. Anthropic says only the user can mark a task complete, even when Duet suggests it is done.
Proactivity Is a Schedule, Not Constant Awareness
In Peter Yang's early-access demo, a brain dump became several trackable tasks and Duet offered next steps. He also inspected a default check-in routine that ran every three hours during his day, producing reminders from connected email and calendar. He thought that cadence was too much and preferred fewer check-ins. Anthropic's launch article says Duet checks connected accounts every few hours, not continuously, and users can tell it what they want to hear about and how often.
The practical value is a notification that leads to a reviewable task, not a stream of ambient suggestions. For a first setup, choose one check-in window and a narrow question: “Which client deadlines need my decision today?” Watch for missed items, duplicates and false urgency. Scheduled background work also counts against the shared Claude plan limits, so reducing noisy check-ins may save both attention and allowance.
Dashboards and Editable Artifacts Are the Standout
The episode highlights Peter's request for a YouTube performance dashboard. He supplied his channel, and Duet assembled charts and content observations from public video information; it was not shown reading his private YouTube Studio analytics in that example. Peter also used a built-in morning-brief skill and showed a recap routine that resurfaced unfinished artifacts. These are strong demonstrations of packaging scattered information into something a person can inspect, but any derived trend still depends on complete, correct source data.
Later, the hosts react to Riley Brown commenting on a slide deck, adding an image and dictating a batch of slide edits. Anthropic's artifact guide supports editing documents, decks and dashboards directly or with Claude. Voice in the shown workflow captured a request and applied it after submission; it was not live, word-by-word editing while Riley spoke. That distinction matters if your work depends on immediate visual feedback.
What the Drip-to-Dashboard Demo Actually Proves
Brock Mesarich wanted email-campaign metrics from Drip, which he could not find as a direct Claude connector in the demo. He connected Drip through Zapier MCP, asked Duet for his latest campaign, and received a dashboard showing delivery, opens, clicks and unsubscribes. It is a useful bridge from an existing business app to an artifact, and a more concrete business test than asking a general agent what it can do.
It is not an unrestricted 9,000-app permission slip. Zapier says the configured connector can expose 9,000+ apps and 40,000+ actions, but each available action depends on the app connection and chosen tool permissions. Its MCP documentation says each successful tool call consumes two Zapier tasks from the existing plan. A marketing connector that can read metrics may also expose subscriber-tagging or campaign actions; choose only the read tools needed for a report and inspect the result against the source account. That also limits the damage from an incorrect request or compromised external content.
The Missing Pieces Are Specific
| Friction | What the evidence says | Why it matters |
|---|---|---|
| Browser handoff | Peter's Uber request stopped at a human-verification challenge. Anthropic confirms that users cannot watch or take over Duet's cloud browser. | Some sites require a person to complete a step; Duet can reach that point but cannot hand over the browser. |
| Desktop | Duet has iOS and Android apps and a web interface at claude.ai, but no native Duet desktop app or access to local files. Claude Desktop and Cowork are separate products with different computer-use features. | Phone-first delegation may not fit a job that depends on local documents or a visible computer session. |
| Team use | The preview is limited to individual US Pro and Max subscribers; Team and Enterprise are not supported yet. Duet channels share one personal agent. | A shared Slack channel or multi-person workflow needs another product and its own permission model. |
| Check-in timing | Duet checks accounts at intervals and may misjudge urgency, according to Anthropic. | A timely business alert cannot be guaranteed just because the assistant is described as always on. |
Andrew sometimes calls the browser issue “no computer.” That is shorthand for the missing accessible computer, not the actual architecture. Anthropic says Duet runs on a cloud virtual machine with a browser; the preview simply does not expose a takeover path when the browser needs human help. Similarly, “no desktop app” does not mean “no access from a desktop browser.” Claude Cowork's local computer use should not be attributed to Duet.
Can You Try Claude Duet Today?
Anthropic currently requires a US Claude account, a Pro or Max subscription and an invitation from the official Duet waitlist. The preview is included in those plans at no separate fee, but morning briefs, check-ins and long tasks use the same Claude allowance. Team and Enterprise access is planned, not live. There is no announced Portugal date in the official help article. For a reader here, the honest next step is to study the demos and prepare a low-risk test, not to assume the app is available locally.
The Original Demos Behind the Discussion
Andrew and Andrew quote several other reviewers. These two hands-on videos show the distinct points the episode leans on:
Peter Yang: tasks, public-data dashboard and the Uber blocker
Peter's full walkthrough provides the context for the task, check-in and failed ride examples.
Brock Mesarich: channels and the Drip integration
Brock's hands-on follow-up shows the channel and Zapier-to-Drip workflow. Claire Vo's independent review explains why the same polished interface still did not become her daily agent.
The First Test Worth Running
If you receive access, give Duet one bounded, reversible job: summarize tomorrow's calendar conflicts and prepare a draft follow-up for one connected inbox. Put it in a dedicated channel, specify one daily check-in, and keep sending disabled until you review each draft. After a week, compare useful alerts, missed deadlines, false alarms, review time and plan usage. Then try one dashboard from a source you can independently inspect. A personal agent earns a place in your workflow by reducing real work, not by winning a feature checklist.
Video Chapters by Topic
The timestamps follow the supplied chapter list and link to the relevant moment in Andrew and Andrew's episode.
The agent model
| Time | Topic |
|---|---|
| 00:00 | Can Duet replace other personal agents? |
| 00:54 | What Anthropic built |
| 02:24 | Channels for separate work |
| 03:09 | One agent or specialist agents? |
| 04:21 | The multiplayer gap |
Tasks and follow-through
| Time | Topic |
|---|---|
| 06:09 | Turn a brain dump into tasks |
| 07:03 | The task interface |
| 08:06 | Proactive recommendations |
| 09:00 | When alerts become too much |
| 10:03 | Desktop and browser limits |
| 11:24 | Uber booking hits human verification |
Outputs and integrations
| Time | Topic |
|---|---|
| 13:39 | YouTube dashboard from public data |
| 15:27 | Skills and the morning brief |
| 17:06 | Scheduled check-ins |
| 18:18 | Drip dashboard through Zapier MCP |
| 21:36 | Task states and review |
| 22:57 | Commenting on artifacts |
| 24:27 | Voice-led slide edits |
Verdict and next steps
| Time | Topic |
|---|---|
| 27:18 | Channels, tasks and threads get confusing |
| 28:12 | Recap unfinished artifacts |
| 29:24 | The hosts' verdict |
Sources and Limits
Primary episode: The Next New Thing, “The Best Parts of Claude Duet and What's Still Missing”, with the supplied transcript. Referenced demonstrations: Peter Yang, Brock Mesarich and Claire Vo. Official checks: Anthropic's launch article, Duet help page, safety explanation and artifact guide; Zapier's MCP documentation. Checked 27 October 2026. The preview can change; the reported demos do not establish general reliability, cost savings or a winner among competing agents.