See pricing
Claude Duet is beautiful. I still forgot to use it
How I AIAI

Claude Duet is beautiful. I still forgot to use it

Claire Vo likes almost everything about Claude Duet's design, especially the onboarding, task lists, artifacts, and channel model. The useful lesson is why that still was not enough: AI agents win or lose on speed, re-engagement, desktop presence, and team fit, not polish alone.

October 21, 2026 · 16 min listen · 13 min read
0:00
–:––

Context

Claire Vo reviews Claude Duet, Anthropic's new phone-based personal agent, through the lens that matters most for AI products: not whether the demo is polished, but whether the agent becomes part of real work. She tests it on inbox and Slack triage, sales pipeline review, a Google Docs sales deck, a launch video workflow, and even buying Warriors tickets. The episode matters to PMs because it separates first-run delight from durable product value, and it shows how latency, notifications, device support, and collaboration shape adoption more than launch polish does.

The Big Idea

Claude Duet proves that an AI agent can be beautifully designed and still fail to become a habit. In a market full of personal agents, the winner is not the agent with the cleanest onboarding. It is the one that closes loops fast, pulls you back at the right moment, and fits the surfaces and team workflows where work already happens.

Claire's review is positive on design and negative on stickiness. She kept seeing good ideas in the product, but still forgot to use it because it was too quiet, too slow, and too isolated from desktop and team use.

Key Insights

Polish solved first-use friction

  • Claude Duet made a strong first impression. Claire said the onboarding was clear about what Duet could do, what it had access to, and how to get started.
  • She described the app as exceptionally polished, with no obvious sharp edges and a design that felt trustworthy from the start.
  • That trust mattered because the product touches sensitive work surfaces like email, Slack, and CRM data.
  • Claire contrasted it with OpenAI's Dots and said Duet felt approximately 100 times more polished at launch.
  • Product lesson: in agent products, good onboarding and visual polish still matter. They reduce fear and setup friction. They just do not guarantee repeat usage.

Primitives made work visible

Claire liked that Duet was not just a blank chat box. It had a set of interaction primitives that made agent work easier to inspect and organize:

  • Connectors and plugins: Duet inherited the plugins already connected to Claude, such as Gmail, Google tools, and Figma.
  • Task lists: Duet showed what goals and tasks were in progress. Claire saw this as a strong pattern because it makes agent work legible instead of mysterious.
  • Artifacts: outputs such as designs, slides, and HTML artifacts were treated as first-class objects, not buried chat messages. Duet could also read artifacts that already existed in her Claude account.
  • Channels: Anthropic borrowed the Slack idea of channels for sub-workspaces. Claire created a sales channel so one thread did not have to hold every kind of work.

The PM takeaway is that agent UX needs visible state. Users need to see what the agent is doing, what it produced, and where a piece of work lives.

Proactivity means re-engagement

  • Claude Duet was pitched as proactive, but Claire's lived experience was the opposite. After setup and a few initial tasks, it never meaningfully pulled her back in.
  • She was not sure whether notifications were off or whether the agent was intentionally quiet. Either way, the product effect was the same. She forgot it existed.
  • Claire argues that generic morning briefings are no longer enough. When several agents can all send a summary, a summary stops being a differentiated reason to return.
  • This is a product growth problem, not just a feature problem. A calm, elegant agent may feel nicer in the moment, but it can also disappear from the user's routine.

Latency broke the coworker feel

Claire's strongest critique was not about intelligence. It was about waiting.

  • She described Claude models as slow enough that she disengaged while tasks were still running.
  • In one example, Duet was shopping for Warriors tickets while she was in the car. The task took long enough that she mentally moved on and forgot to come back.
  • She said some spawned threads took 10+ minutes to finish, return, or ask for permissions.
  • That matters because a personal agent is supposed to feel like a colleague who is always available. If that colleague responds too slowly, users stop turning to them, even if the eventual output is somewhat better.
  • Claire explicitly said she would often prefer a faster agent that did a job 10% worse if it kept her momentum.

For PMs, the lesson is that latency is not just an infrastructure metric. It directly shapes whether users will work with an agent synchronously.

Desktop and sharing were missing

Two missing capabilities kept Duet out of Claire's actual work stack.

  • No desktop app: Claire said she does not want to live on her phone during the workday. She wants to start tasks on mobile and continue them when she is back at a computer.
  • Weak shared context: when she tried to create a launch video using workflows she had built elsewhere in Claude, Duet did not automatically ingest those local skills. She would need to push them as plugins instead.
  • No real multiplayer model: Duet's channels looked a bit like Slack, but they did not provide the sharing and team coordination value that makes Slack useful in the first place.

This matters even for small business users. Claire says that a "personal" agent aimed at entrepreneurs still has to fit a real team context, even if that team is only five people.

Real tasks felt mostly fine

Claire's hands-on tests showed a product that was capable, but rarely the best choice.

  • Inbox and Slack triage: Duet did an okay job, but the priorities and context did not feel fully right.
  • Sales pipeline review: existing connections to Slack, email, and CRM data helped it pull relevant information quickly.
  • Google Docs deck: the output was serviceable for many business users, but it missed Claire's design bar because it did not use Figma or closely match brand details like fonts.
  • Launch video creation: Duet could not easily reuse the specific local skills she had already built in Claude for that workflow.

The product lesson is that broad competence is not enough in a crowded category. A user with several agents will only keep the one that is clearly better on their highest-frequency, highest-value jobs.

Mental Models & Frameworks

Attention is the bottleneck

Claire frames personal agent competition as an attention market, not just a capability market.

  • A new agent is not competing against no solution. It is competing against every other agent the user already has.
  • In that environment, a product must win three things at once: initial trust, repeat recall, and enough value per interaction to stay in the stack.
  • Use this model when prioritizing roadmap work. If a feature improves demos but does not improve recall or repeated use, it may not move the real outcome.

Agent primitives need structure

Claude Duet worked best when it gave the agent a visible operating model:

  • connectors for access,
  • task lists for progress,
  • artifacts for outputs,
  • channels for separation of work.

This is a useful framework for PMs designing agent products. A good agent surface usually needs all four: what the agent can reach, what it is doing, what it produced, and where a workstream lives.

Synchronous work needs speed

Claire's review implies a simple split between two kinds of agent work:

  • Synchronous tasks: work where the user is waiting in real time, such as inbox triage, shopping, or interactive document edits.
  • Asynchronous tasks: work where the user can leave and come back later, such as background research or overnight summaries.

Use this model when deciding model choice and interface behavior. Slower reasoning can be acceptable in asynchronous tasks if the loop reliably closes. It is far more damaging in synchronous tasks because delay breaks the sense of collaboration.

Personal agents face team gravity

Many AI products start as solo tools, but work naturally pulls toward collaboration.

  • Claire sees Claude Duet as a strong single-player product.
  • Her actual work, even as a small business owner, still involves several collaborators.
  • That creates "team gravity." A tool that cannot be shared, handed off, or seen by others has a lower ceiling in work contexts.

Use this model when scoping product strategy. Decide early whether the product is intentionally solo or whether collaboration is part of the core job to be done.

Decision Principles

Principle: Optimize day-two pull

  • When: building or prioritizing a personal agent that already demos well in onboarding.
  • Why: Claire liked Duet on first use, then forgot to return because the product did not create a strong reason to come back.

Principle: Match speed to mode

  • When: choosing models, task routing, or interface patterns for agent work.
  • Why: a slower, stronger model may be acceptable for background tasks, but Claire's review shows that synchronous work lives or dies on responsiveness.

Principle: Choose primary surface early

  • When: deciding whether an agent experience should be mobile-first, desktop-first, or genuinely cross-surface.
  • Why: Claire's usage broke because the product lived on her phone while her real workday lived on desktop and in meetings.

Trade-offs & Nuance

Calm design versus recall

Claire clearly liked that Duet was subtle, tasteful, and not noisy. The trade-off is that subtle products are easier to forget, especially in a category where users already have multiple agents competing for attention. A calmer agent may feel better in brand terms while performing worse on habit formation.

Better reasoning versus speed

Claire said she likes Opus 5.5 and thinks the models do strong work. Her critique is that model quality does not excuse visible delay. For many work tasks, users would rather have a result that is slightly worse but fast enough to keep momentum.

Mobile convenience versus desktop flow

Duet's phone experience felt zippy and engaging, and Claire liked being able to kick off tasks on the go. The trade-off is that many work tasks still need follow-through at a desktop. Without that continuation path, mobile convenience becomes a partial solution.

Personal focus versus team value

A single-player agent is simpler to design. It avoids permissions, collaboration complexity, and shared-state problems. The cost is lower usefulness for small teams, which is exactly where many busy professionals and small business owners actually operate.

Common Mistakes

Mistake: Confusing polish with retention

A clean first-run experience can hide a deeper product weakness. Claire's review shows that a beautiful onboarding flow and elegant design can still produce weak retention if the product does not create a recurring use habit.

Mistake: Calling silence proactive

An agent is not meaningfully proactive just because it can perform actions in the background. If the user is not reminded, nudged, or looped back in at the right moment, the product will not feel proactive in practice.

Mistake: Ignoring team workflows

It is easy to design a personal agent around one user's chat history and tools. Claire's experience shows that this breaks down when the real job involves a small team, shared decisions, or handoff into collaboration tools like Slack.

Practical Application

Audit the second session

Run a simple product review on any agent feature after the first task is complete.

  • Check what happens in the next 24 to 48 hours.
  • Note whether the agent sends a useful nudge tied to unfinished work, changing context, or a new opportunity.
  • Treat second-session return as a separate product problem from first-session onboarding.

Track loop-closing latency

Measure more than model response time.

  • Time the gap from user request to first meaningful update.
  • Time the gap from request to usable output.
  • Time the gap from request to permission request, if the workflow requires approval.
  • Review the tasks that exceed a user's likely attention window, because those are the tasks users will forget.

Test full workflow chains

Do not evaluate an agent on toy prompts alone. Use end-to-end workflows like Claire did.

  • Ask the agent to pull from inbox, Slack, and CRM data.
  • Ask it to create a document or deck from that context.
  • Score output quality, prioritization quality, permission friction, and whether the user stayed engaged long enough to finish the loop.

Make state visibly persistent

Claire liked Duet's task lists, artifacts, and channels even when the product fell short elsewhere. That is a useful build pattern.

  • Keep goals visible.
  • Keep outputs easy to inspect.
  • Keep workstreams separated by project or job.

Users trust agents more when they can see progress and find the resulting work later.

Decide sharing before scale

If the target user is a founder, PM, or small business owner, decide early whether the product is purely personal or intended to work with a small team.

  • If the product is personal, optimize hard for solo speed and simplicity.
  • If the product needs to enter team workflows, design sharing, handoff, and shared context deliberately instead of adding them later.

Questions to Consider

  • Where is your AI product optimized for first-run delight but weak on the user's second or third visit?
  • For each agent task in your product, what is the maximum wait time before a user is likely to abandon the interaction or forget the task exists?
  • If your agent starts on mobile, what breaks when the user needs to continue the same work on desktop?
  • Are your target users truly working alone, or are they trying to coordinate with a small team even when the product is framed as a personal assistant?

Bottom Line

Claude Duet shows that great agent design is necessary but not sufficient. Claire Vo's review suggests that the real bar for a personal work agent is higher: it must be fast enough to feel collaborative, proactive enough to be remembered, and flexible enough to live across desktop and team workflows, not just on a polished phone app.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
Claude DuetAnthropic's mobile personal agentThe main product under review, tested across inbox, Slack, CRM, docs, and shopping tasks.
Claude CoworkAnthropic's desktop Claude work surfaceClaire uses it as the comparison for the kinds of work someone might want to continue from Duet.
Meta MusePersonal agent productA benchmark for agent primitives like task tracking and side workspaces, and one of the few agents Claire thinks has stronger re-engagement.
Grok BotAgent product from xAIA benchmark for desktop presence, visibility, and team-agent usage in Slack.
DotsOpenAI's personal agent experienceUsed as a contrast point on launch polish, where Claire thought Duet was far stronger.
Google DocsDocument creation and editing toolDuet used it to generate a sales deck from the phone.
SlackTeam communication toolAnthropic's channel concept borrowed from Slack, and Slack remained the missing benchmark for true multiplayer collaboration.
FigmaDesign toolClaire wanted stronger brand and design fidelity in Duet's deck creation workflow.

People to Follow

Claire Vo

Claire Vo hosts How I AI and reviews AI products against real operator workflows instead of demo-only use cases. This episode is a strong example of product critique from a PM lens: what matters after the launch video ends and the real workday begins.

Notable Quotes

"There are only so many three point morning briefings that can pull me back into a product." (Claire Vo)

"I need you to be a little bit more annoying." (Claire Vo)

"The reality is I do not want to be on my phone during the day when I'm doing work and sitting on meetings." (Claire Vo)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free