Claude Duet flagged something in my inbox I had no idea about: my NeverBounce credits had run out. That means my form leads weren’t being email-checked, which is an email reputation problem waiting to happen.
Duet's answer to "what's waiting on me?". I removed the names of the people and the client.
I got off the waitlist Monday night. Duet is Anthropic’s new always-on agent: you hand it tasks, it works in the background, and it comes back when something needs you. It launched October 21 as a research preview, for US Pro and Max subscribers by waitlist (Anthropic’s launch post).
The test: what’s waiting in my inbox
Duet suggested this task itself in its first message, so that’s the one I ran.
It says it scanned 106 of the 109 threads since October 18. Heads up: these numbers are Duet’s own count, not mine. It also gave me a breakdown of what it skipped (own alerts, newsletters, service notices, DMARC reports), but those categories add up to more than the threads it said it scanned. I’m not printing them. It said it missed 3 threads because Gmail search returns at most 50 at a time.
Of the six things, only one had a person waiting on it. A colleague wanted a video embedded on a page, and Duet noticed I was only cc’d, so it “may not be yours”. That’s a good call.
The other five included the NeverBounce credits, a domain misconfigured on Vercel, and a Reddit reply on my Pi 5 post.
What I liked
The app is its own thing. Different look, different feel, different functionality from Claude chat or Claude Code. It reads as aimed at business owners, and Anthropic’s product page says as much: unpaid invoices, payroll prep, more than 6,000 connectors (Claude Duet). To me it’s an attempt to bridge what Claude Code can do and the “no technical experience required” ease of Claude chat into its own thing. It feels simple. There’s no setup to speak of. I’ve only used the web app so far, not the phone app.
Tasks cost nothing to create. A sentence like “check my email every day at 10am” is all it takes to make a recurring task. Heads up, that’s an example of the kind of prompt it handles, not one I ran.
It doesn’t ramble. Replies are short and to the point. Claude Code tends to drag things out. Duet doesn’t.
What was weaker
GitHub. It clearly knew my repos. Its own example used my handle. And it still made me type the repo out as owner/name. That felt like friction by design.
It isn’t as proactive as I hoped. Its first message did put three suggested tasks on my list. What I wanted was a plan: here’s what I can do for you, here’s what I’d recommend, and here’s the schedule. Instead I got suggestions with no order and no cadence.
It doesn’t explain what it can do. When I asked it to write a section for this post on how it differs from Claude chat and Claude Code, it wrote one without any trouble. So it can explain itself. It just doesn’t volunteer it, and a new user can’t tell what’s worth asking for.
I asked Duet to write a "what it is and how it differs" section for this post. This is its answer, and I'm labelling it as Duet's words.
My own always-on agent
I’m building an always-on agent of my own, called Raspi Solo, and I’m only two days in. I wrote about the hardware in the Pi 5 post and about what bit me after three weeks. Solo is the brain on top, and it works a bit differently from what you’d guess.
- It doesn’t do the work itself. It hands jobs to other lead agents.
- It runs on scripts, to keep token cost down. Scripts fetch the data. Solo only reads the results and decides what’s worth bringing to me.
- It can nudge itself. It has access to the scripts and a few ways to wake itself up and check on things.
So Duet’s pitch, an agent that keeps working while you’re away and tells you when something needs you, isn’t new to me.
I’d just taught Solo to check my email, so I pointed it at the same inbox. It brought me the same three items later, NeverBounce included. It had only just learned to read my email, so I wouldn’t read much into the gap.
Where I landed so far
If you already run your own agents, I’d expect you to have a version of this already, and probably a better one tuned to your work. On the inbox task, Duet got there first and my own agent, two days into being built, brought me the same item.
If you don’t, I can see this replacing Claude chat for business use, and making Claude Code feel less far away. That’s a big group with money, and it makes sense that Anthropic is going after it.
The concept is clear, and for that role I think Duet fills it quite well. Checking email is cool. But Claude Code with access granted can do the same thing, so it isn’t ground breaking. What I’d want to see is Duet making the most of itself without me having to know what to ask. It isn’t there yet.
If you want in
Duet is for US Pro and Max subscribers, by waitlist. Team and Enterprise aren’t included yet, and it’s free during the preview. The waitlist is on Anthropic’s Duet page.
Two things you can do while you wait. Duet pulls in your Claude memory and connected apps, so the more you’ve set up there, the more it knows on day one. And decide what you’d hand it first. The inbox is the obvious one.
I’ll write a follow-up once I’ve used it more. Happy to answer questions in the meantime.
Related:
- Claude Design Brand Kit: GitHub Repo to Live Page in a Night: The same kind of post, a launch-week test of a new Anthropic release on real work
- I Turned a Raspberry Pi 5 Into a Claude Code Box. Three Things Broke First.: The hardware behind the always-on agent I’m building
- Three Weeks of Claude Code on a Raspberry Pi 5. The Login Held. My Temp Folder Didn’t.: What running that box unattended actually looked like