实测 Claude Duet：精美设计的个人 Agent 为何难敌 Grok 与 Muse？

I have good news and I have bad news. The good news is Anthropic just released a new personal agent and it's really welldesigned. The bad news is I have personal agent fatigue. I am so tired of new personal agents. And while this one is super super well-designed and I'll go through what I love about it, I just don't need another one. And so to compete for my attention, to compete for being my number one personal agent, you have to be not good, not great. You have to be exceptional. And we're going to see in today's episode if Claude Duet, their new personal agent, is exceptional. Let's get to it.

Okay, before we go into my experience with Claude Duet, just a very, very quick overview of what this is. It is a phone, iPhone, Android based personal agent that is synced up with Claude. So, it is proactive, although I will go through how proactive it really is. It's proactive. It can use all the tools and plugins that your cloud has access to. And they really position this as being a personal agent for small business owners, for busy people, for professionals. So, none of this like your mom's going to use it with Muse and none of this like you have to be a developer. This is like solid middle of the road. Business user, SMB, entrepreneur, busy PM, busy CEO, um, agent. The things that you would reach for in co-work on your desktop, maybe you'll start reaching for on your phone.

Now, you all know that I really love OpenAI products. I use codecs all day, every day. But I will say one thing that Anthropic did so so so well that OpenAI's Dots did so so so poorly is the polish of the agent design. This app is beautiful. It is welldesigned. There are no sharp edges. It is clearly incredibly well thought through from a design perspective. Everything from the very very early test flight versions that I got access to to the current beta that's out in research preview. From the second you onboard, they are really thoughtful about walking you through what to do. It looks really nice. It feels really trusted. I had zero problems getting up and running out the gate. It was a really nice onboarding. It was a little similar to my experience with Muse where it was very clear upfront what to do.

The product itself has some primitives again like Muse that I think you're going to see over time emerge as like core for agent design. Of course, it has plugins and connectors, so it inherits all your plugins from Claude. So, if you have your Gmail set up, your Google setup, your Figma setup, whatever you have set up in Claude, it's immediately available here in Duet. It also has task lists, which is something I really liked from Muse's agent design. I like the idea of keeping track of what goals and tasks I'm doing and what goals and tasks the agent is doing. So, I think the task list view is something that's really nice here and something that was really thoughtful.

And then what it has that again Muse has, but it does just a little bit more polished, is it treats artifacts as first class citizens. So if you're used to artifacts in claude code or in co-work if you're generating HTML artifacts designs slides anthropic has made those first class citizens in duet so that whenever you're generating stuff in your little chat with your little duet it's going to create those artifacts and it actually already can read the artifacts existing on your cloud account. So when I showed up here and looked at my clot account, it already had all those artifacts that I had generated over time in there and I had a really good sense of what I could do next.

Finally, the last thing I really love about Duet is the idea of channels. So we're all trying to decide what the primitives of are for agents. In Codeex, you have all these threads in chat. You have all these chats. In Claude, you have threads and chats and projects, all sorts of things. In Muse, you have your main chat and side chats. Everybody's trying to figure this out. And Anthropic borrowed a little bit from Slack, which I feel like is everybody's favorite tool to borrow UI designs from. So now you can kind of think of your duet as a main thread, a main chat, but but also side channels that you set up. Like here, I created a sales channel for working on sales projects.

And then I'll say my last little thing that I really like that I did give the anthropic team credit for, their little dot icon is really cute. It's abstract. It's definitely anthropic claude e. It's cute, but I have no problem with this little, dare I say it, dot. Um, it's just adorable, not offensive, not weird. I like it. Yes, of course. I love my personalized Muse consumer agent and that engenders a lot of joy with all the little characters. Of course, I love my little Grock bots and their little icons, but this is inoffensive and I think it's just fine. And as somebody who cares about design, I thought it was pretty good.

Now, what are some things that I did not like about Duet? Well, I will tell you for a proactive agent, it just wasn't that proactive. I don't know if I forgot to turn on notifications or it's just not that annoying, but once I set up Duet a couple days ago and did a first couple tasks, it never pushed or bothered me to do anything ever again. It also doesn't show up in the desktop app. So, it's a really disconnected experience. And so, while I felt like the initial experience was pretty good, I ended up not using it a ton. one, because I have a million other personal agents, and two, it really didn't push me to come back.

I think one of the things that these agent designers are really going to have to figure out is what is a notification strategy that actually engages users. I get them from DOT all the time. They're not that exciting. I get them from now duet. They're not that exciting. Like, there's got to be something better than the morning briefing to engage agent users on why they would come back to the agent over time. And right now, I don't think any of these agents, except for maybe Grockbot and Muse, have really figured this out. So, it's not as proactive as I want. The notifications aren't noisy enough for me to remember it. You know, I've said it's elegant and well-designed and subtle and not annoying, and those are all positives in some ways, but from a growth hacking perspective, and the ability to get me to re-engage with your product, that doesn't work. I need you to be a little bit more annoying. I need you to be a little bit more proactive. I need you to be a little bit more noisy or I'm truly going to forget you in the ecosystem of all my agents.

So, what were some things that I asked it to do and how well did it do it? Well, the first thing I asked it to do was just this like basic kind of like go through my email, go through my Slack, figure out what's going on. I would say it didn't do an exceptional job at that. It did okay. It came up with a couple things, but the priority didn't seem right. The context didn't seem right. So, I do think there's a little work on the quality of the interaction.

I asked it to find me good Warriors tickets. You know, go Dubs. I want to go see the Warriors when they play the Pacers. It's a noon game. I can actually take my kids. And so, I've had all my agents try to shop for tickets for me, shop for shoes for me. I thought this was a good eval. I had it go shop for me. And this is the second thing that I think is bad about it. So, the first thing that I didn't love about Duet is that it is not that proactive. The second thing that I don't love about it, and this has been my experience with Claude generally, is it's slow, man. And I get so disengaged waiting for the models to finish. The killer feature of Muse Meta is that it is fast. In fact, Codex with their GPT ultraast models, it just keeps me engaged because I am always moving to the next step, always getting something done. And so even though I really love Opus 55 as a model and even though it does really good work, the performance is so slow and you really feel that in a personal agent and it degrades my desire to work synchronously with an agent. You know, one of the benefits I think of an AI agent is that it is like a colleague, but a colleague who's up 24 hours a day and responds to immediately. But if that colleague is slow, if that colleague goes to a meeting to get work done, like you kind of forget and don't want to go to that colleague over your more responsive ones, even if the more responsive ones do like 10% worse of a job.

And so it was shopping for these warrior tickets for me as I was like passenger princess and driving to dinner. And like I kind of forgot that it did it cuz it took so long. It didn't close the loop. And then by the time that I came back, I was off to something else. And so even though it tracked it in this, you know, checklist of things that it was doing, the performance was really frustrating. And I think this adds up, especially if you use a single chat, single agent, where you're firing off a bunch of tasks. Like you're saying, "Okay, check my email and I need you to make a deck and I need you to make edits on this thing, and let's figure out what we do for social." And in Claw Duet, it those all kicked off individual threads, but then they took so long and I'm telling like 10 plus minutes to finish those threads or come back to me or ask for permissions. And because it wasn't noisy, because it wasn't, you know, like in my face, I actually forgot about all those threads. And so while like foundationally it's really welldesigned, the models are very good, the intelligence is there, and like I can't knock anything specific about the user experience. The reality is like the latency and the inability to close loops meant that this particular agent product just really fell to the bottom of my use list.

I think the other thing that they are going to have to solve ASAP. And again, this is in research preview. So, it's an early access list. It's in beta. I'm sure they will get there is a desktop app. The reality is I do not want to be on my phone during the day when I'm doing work and sitting on meetings. I need a desktop app to engage with my agents. I think that's why like Grockbot does so well is there's a desktop agent. It's cute little icon. It's sitting there staring me in the face. I'm always chatting with it like it's Slack. Same with Muse. You know, I started using it as a mobile app, but very quickly download the desktop app. And then even same with Dots from Codeex, which I don't even use, but at least it's in the desktop app. Like at least my dot bay is staring me at the face in Codeex like with a little icon at the bottom.

And so what I would say about Duet is like it's got all the makings of a great agent, but it needs to show up. Maybe that's intentional. Maybe the design is to be calm and soothing and helpful but in the background. But I do think to compete for attention right now, it is not enough to have beautiful design. It is not enough to have these artifacts which is actually really awesome and one of the best parts of the Claude ecosystem. The agent itself has to be super proactive. The agent itself has to be super fast and it has to get stuff done quickly enough so that I don't disengage with it.

Okay, so that was a very long aside for buying Warriors tickets. A couple other things that I had it do, good, bad, left, right, you know, standard evals. I had it look at my sales pipeline. Nice things there. It was already connected to Slack. It was already connected to my email. It was already connected to our CRM. So, it was able to pull in that data pretty quickly. Um, I had it create a deck in Google Docs. I have a really high bar for design and so I was disappointed that it didn't use pitch or gamma or whatever, but it worked. It was created in Google Docs and it was kicked off from my phone. I think for like the kind of everyday business person who maybe is not as particular as I am about design, having an agent kick off a deck on your phone on the go with Duet. So, I think that was like good, not great, but I have a really high bar for that stuff.

And then one other point of like friction and disappointment, again, this is a minor thing, but I think for heavy users this matters. I've been making all these launch videos in Claude. I've been spending a whole bunch of time doing like chat PRD launch videos in Claude. They're super cool. Um, and I wanted to see if I could create those in Duet on my phone. The reality is unfortunately duet does not ingest all your local skills. Um so I probably need to push these skills as plugins to claude and then it can ingest them into Duet. But again I think with all these product surface areas and this is not just an anthropic problem. This is an open a problem. um is like the shared context, the shared experience, what ports over elegantly, what does not port over elegantly, I think is going to be an open question for folks who want to really migrate to Claude Duet.

I'd say the final thing that's really missing is look, we don't work alone and Duet right now is clearly a agent for single player use. In fact, like most of the Claude projects feel like very good products for single player use. I just don't think even with claude tag claude and slack all these things I don't think they've quite cracked multiplayer and I've been moving more of my grobbot usage to team agents and deploying those in Slack. So while Claude Duet with these channels might feel like Slack, it doesn't have a multi-player capability. Like I'm not running a team of 500 anymore. I do not need like the full enterprise deployment, but I am a small business owner. The problem is I have a small team, so you know, I have like five humans that I interact with on a daily basis. If I don't have a way to share my agents with them, I'm not sure if it's going to fit into my stack.

So, that is my very quick preview, early review, early thoughts on Claude Duet, the new personal agent from Anthropic. I would definitely give it a try. Again, I maybe have different use cases, different preferences, different desires than other folks. I think it is approximately 100 times more polished than Dots was when it came out. So, I think that's really nice. I think it is quite elegant. It does the job well. I would just say there's like fit and finish things that they really need to work on. So, give it a try. Sign up for early access if you can. Let me know what you think and let me know if Claude Duet becomes part of your stack because I'm just not sure it's going to become part of mine. Thanks for joining How I AI. Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review which will help others find the show. You can see all our episodes and learn more about the show at howiipod.com. See you next time.