A few weeks ago, Claude started beta testing a new feature called Projects (not to be confused with a previous feature…also called Projects).
Projects has been nothing short of transformational for my workflow. It’s essentially one large chat canvas. For each project that you’re working on, you feed it a bunch of random thoughts and assignments, and it spins up individual threads to handle them.
I’ll go on record and say that Projects is one of the top three new Claude features of 2026.
However, the introduction of Projects has left a newly uncovered product hole. If Projects can spin up individual threads to handle discrete tasks, how are users supposed to coordinate inputting work into Projects?
Right now, the answer is: click in and out of each Project, one at a time.
But that doesn’t really gel with how I want to work with Claude, especially as we round the bend towards new agent-first interactions with AI.
What I Want from Duet (Or Any Agent)
What I want is a top-level coordinator for my Projects. A chief of staff, as it were. An agent. If a Project can push a task to a thread, I want an agent who can push a task to a Project.
About two weeks ago, Claude announced Duet, and I thought my wishes had been granted. But two weeks of testing Duet has left me unsatisfied.
In two weeks with Duet, I haven’t had a single moment where I walked away thinking, “Wow, that was really incredible.”
Here’s my dream scenario: I’d go for a walk, put Duet into voice mode, and in one session we’d review tasks due today in Todoist, loose Project tasks, and emails that require responses.
During this chat, Duet would know everything I know. It would have access to all of my Claude memory, Project memory, my emails, MCP connectors, and more. I’d never feel like I know more than Duet.
As we settle on a plan, Duet would send tasks to the corresponding Project, and that Project would spin up, or resume, threads to handle them.
At all times, I could jump between Duet and the Project itself, and both would stay fully up to date.
But, two weeks in, Duet hasn’t nailed this.
Duet, The Anxious Intern
And it’s not for lack of trying. Duet wants so eagerly to help me that it raises its hand to pitch in at every opportunity.
The problem is that Duet does so without first understanding the larger context. It lacks the memory, skills, and history of everything I’ve been working on in the main Claude app.
For example, a client emailed me this morning, and Duet immediately pinged me, asking something to the effect of, “Hey, you got this email from a client who has a problem. Want me to start fixing it by doing this, such-and-such?”
However, Duet didn’t bother to first cross-reference this client’s email with an entire conversation I had with the corresponding Claude Project about exactly how to handle it.
The right answer was sitting in the Claude Project, but Duet took my client’s email at face value and immediately was like, “Here’s how I can help,” without actually thinking through just how to help.
Right now, it feels like every time Duet tries to answer an email for me, it does so with the same competency as an overeager intern on their first day. There’s no context, and no awareness of the existing playbooks for how I actually want work done.
Additional Paper Cuts
And while Duet’s biggest crime is acting without context, there are additional rough patches throughout the experience.
Constant permission prompts, so it can’t truly work autonomously.
Buggy website sign-in experience. I haven’t yet completed a sign-in.
Voice mode disconnected from main chat. Always “sending” back to chat.
Proactive email drafts miss what the sender really wanted from me. The subtext.
A Fine Line Between Helpful And Interrupting
Finally, a good assistant would know when to interrupt me.
Like anyone in a client-service business, I sometimes don’t want to know something, or at least not yet. And sometimes someone calls with an “emergency” that is anything but.
A human assistant would know my preferences, when to interrupt and when to let something wait a beat.
And if I’m already getting a notification about a new email, does Duet really need to ping me too?
Too many times in the last two weeks, it’s felt like Duet has interrupted me at the wrong times.
Some of this may come down to training Duet on preferences more nuanced than a simple pause-notifications toggle.
Final Thoughts
If Projects represents the best of what Claude can be, Duet represents an unfinished product without a central operating thesis.
I’m not trying to be critical to be mean. I’m voicing these concerns because I think Duet could be the missing layer that sits on top of, or beside, Projects.
The test will be: can I go for a walk, start a voice chat with Duet, and talk through my day, my inbox, and my Todoist list, and have Duet push the right tasks to the right places with all the right context, so Projects can get to work without asking me to tap a permissions button and just come back to me when everything is ready?
Here’s hoping.