Claude Duet Is Beautiful. That Doesn’t Make It My Go-To Agent.

Claude Duet kept track of the work I gave it. I still forgot about the work.

That is the tension in Anthropic’s new personal agent. It has thoughtful onboarding, task lists, artifacts, and a genuinely polished mobile interface. But when I sent it off to find Warriors tickets, the task took long enough—and the follow-up was quiet enough—that I moved on. A checklist showing that an agent is working is useful. It is not the same as getting me to a decision.

I have personal agent fatigue. I use enough of these products that a new one has to earn its place against tools already embedded in my day. Being well designed gets my attention. Closing loops is what keeps it.

This was an early-access test of Claude Duet, while the product was in research preview. My experience is a snapshot, not a claim that its limitations are permanent. But the test surfaced a practical question: can a beautifully designed agent become a reliable coworker if I keep forgetting to return to it?

The interface gets the important foundations right

Duet is a phone-based personal agent for iPhone and Android, connected to Claude and positioned for small business owners and busy professionals. Think of the work you might reach for Claude Cowork to handle on your desktop, initiated from your phone instead.

The onboarding was excellent. It explained what Duet could do, what it could access, and how to get started. I had zero problems getting up and running. From the early TestFlight versions I used, the product felt unusually considered.

More importantly, it did not ask me to rebuild every connection. Duet inherits plugins and connectors from Claude. If your Claude setup already includes Gmail, Google, or Figma, those connections are available to the agent. That makes the starting point much more useful than an empty chat box.

Three interface choices stood out in actual use:

• Task lists made the agent’s work visible. I like being able to distinguish what I am doing from what the agent is doing.
• Artifacts had a dedicated place in the experience. Duet could already read artifacts on my Claude account, so I arrived with existing work available rather than an entirely blank workspace.
• Channels gave separate projects somewhere to live. I created a sales channel rather than putting every request into the main conversation.

The channels borrow a useful idea from Slack: a main conversation can coexist with focused spaces for particular work. Agent products are still sorting out whether their organizing unit should be a chat, thread, project, or something else. Channels were an understandable choice here.

The catch is that good organization solves only part of the problem. I could see where the work belonged. That did not guarantee I would come back when it was ready.

Inbox triage and ticket shopping exposed the follow-through problem

My first request was basic: go through my email and Slack and figure out what was happening.

Duet produced a few things, but the priorities and context did not seem right. It was okay, rather than exceptional. This is an important distinction for anyone evaluating an assistant: finding activity is easier than understanding which activity deserves your attention.

I did not have a precise prompt or a scored benchmark to share from this test. The result was my judgment as the person who knew the underlying work. If you adapt this evaluation, compare the agent’s proposed priorities with your own before letting it determine your day. That is a suggested testing method, not something Duet demonstrated automatically.

The more revealing task was shopping for Warriors tickets. I wanted to see the Warriors play the Pacers at a noon game, which meant I could take my kids. I have been asking my agents to shop for tickets and shoes, so this was a familiar, practical evaluation.

I kicked off the request while I was a passenger on the way to dinner. Then I waited. Eventually, I forgot about it. By the time I returned, my attention had shifted.

There is no completed ticket purchase to report here. The meaningful outcome was the broken interaction: Duet tracked the task, but it did not bring me back to resolve it.

That pattern became more noticeable when I gave it several requests. It could kick off individual threads for email, a deck, edits, or social work, but those threads could take 10-plus minutes to finish, return, or ask for permissions. Because the agent was quiet, I forgot about the open threads.

I cannot tell you whether I had missed a notification setting. I explicitly wondered about that during the review. What I can say is that after setting it up and trying the first few tasks, I did not experience the kind of useful prompting that would pull me back in.

A morning briefing does not solve this for me. I already get briefings from other agents. The useful interruption is more specific: the work is ready, a permission is blocking progress, or there is a decision I need to make. That is the behavior I wanted, not an observed Duet capability in this test.

Speed matters more when the agent is your conversation partner

I liked Duet’s interface responsiveness. The phone experience itself felt zippy. The work behind the interface was the slow part.

That distinction matters. A polished app can still feel slow if every meaningful exchange requires a long wait for the model or agent to finish.

I use Codex throughout the day, and one reason I stay engaged is that I can keep moving to the next step. Speed is also the feature I especially appreciate in Meta Muse. A responsive agent makes it easier to stay in a working rhythm.

With Duet, the delay reduced my willingness to work synchronously. An agent can be available around the clock and still feel unavailable at the moment I need it. I may choose the more responsive option even if I think it does a somewhat worse job, because it keeps the work moving.

This is not a measured head-to-head model comparison. It is a description of what made me reach for one product over another. Intelligence, latency, and follow-through combine into the experience of reliability.

The missing desktop experience compounded that problem. At the time of my test, Duet did not show up in the desktop app. I do not want to spend my working day on my phone while I am at my computer or in meetings.

The handoff I want is straightforward: start something asynchronously on my phone, then finish it at my desk. Muse’s desktop app supports my habit of returning to that agent. Grok Bot also stays present in my working environment. Duet’s mobile-only experience left it disconnected from where I do much of my work.

The lesson for evaluating an agent is to test the return trip. Starting a task from a phone is only half the workflow. Where will you review the result, answer a question, and approve the next step?

Sales work was useful; brand and local skills did not carry over cleanly

The business tasks were a better fit for Duet’s existing connections.

When I asked it to look at my sales pipeline, it already had access to Slack, email, and our CRM. It could pull in that data pretty quickly. That was a concrete advantage: I did not need to assemble and paste all the inputs myself.

I also had it create a deck in Google Docs from my phone. The deck was not bad. For an everyday business user who wants to generate a sales deck or document on the go, that could be useful.

My disappointment was design fidelity. I have a high bar, and I wanted it to use Figma and match our design, including fonts. It did not do that. The result met a more basic standard than the one I had in mind.

The copyable part is the connected starting point: use the business systems the agent can already access to produce a draft. The added guidance is to make the output destination and visual requirements explicit, then inspect the result. This test does not establish that a connected Figma account means Duet will choose Figma or reproduce your brand correctly.

The launch-video request exposed a different boundary. I had been making launch videos for ChatPRD in Claude and wanted to create that kind of work from Duet on my phone.

Duet did not ingest all my local skills. The setup behind those videos therefore did not simply travel with me into the mobile product.

I suspected I would need to publish those skills as Claude plugins for Duet to access them. That was a proposed next step, not a fix I completed and verified during the episode.

This is worth checking before moving any established workflow to a new agent surface. Existing artifacts, connected services, and local skills are different kinds of context. In my test, some carried over elegantly; others did not. Access to the same account did not mean access to every ingredient that made the original workflow work.

Channels organize solo work. They do not make an agent a teammate

Duet’s channels felt familiar because they resembled Slack. But they did not provide the multiplayer sharing that makes Slack useful to my team.

I am a small business owner, which should make me a good fit for this product. I am also working regularly with about five humans. I do not need a full enterprise deployment, but I do need ways to share work and agents with those people.

I have been moving more of my Grok Bot usage toward team agents deployed in Slack. Duet, in this test, was clearly a single-player agent. Organizing my own conversations into channels did not remove that limitation.

That changes the buying—or habit-building—decision. A solo professional may care most about access to email, documents, and a convenient mobile interface. A small team may care more about whether an agent’s work can become shared work without one person relaying everything.

My verdict: Duet is worth trying for Claude users who want a polished, phone-first assistant for individual business tasks. The inherited connections, visible task tracking, and accessible artifacts are worth copying. Generating a serviceable business draft from your phone is a real use case.

It did not earn a place as my primary agent in this test. Slow task completion, quiet follow-up, the missing desktop experience, and the lack of team sharing kept it below tools already embedded in my day.

Human judgment still mattered at every consequential point: deciding whether inbox priorities were right, whether a deck met the brand standard, whether a workflow had the skills it needed, and whether the agent actually helped finish the job. The standard is not how elegantly an agent accepts a request. It is whether I can trust the work to come back, make the next decision, and close the loop.