With the advent of Muse, it seems AI is finally starting to make its way to the masses (at least according to my X feed, which is of course not really representative of the masses, but it’s still a somewhat useful barometer).
Ejaaz@cryptopunk7213
Flight delayed 7 hours - asked Muse to file for compensation. 5 mins later i had $250 credit in my delta account. it even found and rebooked me a new flight. it just figured everything out. even responded to the support email itself. shit feels like magic.
11:35 PM · Sep 21, 2026 · 1.3M Views
161 Replies · 279 Reposts · 5.99K Likes
Anand@anandragn
Things my Muse agent got done while I took a nap this afternoon: 1. Cleaned up my ancient Hotmail inbox from 96% full to 60%. Went through 80K emails, deleting 20k+ old promos and newsletters with my approval, keeping a full audit log of everything. 2. Chased down a missing…
10:53 PM · Sep 26, 2026 · 73.6K Views
45 Replies · 12 Reposts · 488 Likes
Joe Devoy@JosephDevoy
Yesterday I uploaded my auto insurance policy to Meta @Muse and asked for a better rate with identical coverage. In ~5 minutes, it found a policy saving me $3,500 /year, bought it, and canceled my old one. I knew I was overpaying. I’d abandoned several attempts to switch because …
9:17 PM · Sep 10, 2026 · 626K Views
55 Replies · 31 Reposts · 549 Likes
What’s interesting to me here is that Muse brings almost nothing new to the table in terms of functionality (the one exception perhaps being the ability to call businesses, though even that was technically doable with a little work). Claude and ChatGPT are both great at computer use, and they can both connect to your email. Those are really the only two things needed to accomplish most of the tasks in the X posts above. And, at least in my experience, they’re better at pretty much everything than Muse.
So what changed? The cute furry little dude whose appearance you can customize probably didn’t hurt. The real answer, though, seems to be the user experience. When you go to chatgpt.com, you get a chat UI with a couple of example suggestions. When you open Muse the first time, it walks you through setup. When you ask it to do something in your email, it holds your hand through the connection process.
Beyond that, much of what’s going on just isn’t exposed to the user. You ask it to do something in your email, and it just does it without giving you approval prompts along the way. If you ask it to check on something daily, it doesn’t explain that it’s creating a scheduled job, it just says okay and does the check. If something prevents it from doing the task one day, it lets you know.
Now that AI agents are really starting to hit the mainstream, I maybe be able to offer some useful advice from my experience.
Muse feels much more like employing a virtual assistant than using a piece of software. You talk to it, it does stuff, and then it tells you what you need to know. That’s the whole thing, and it means it’s easy for people who don’t spend all day playing around with AI to use it. Compare that to a product that makes you pick a model and a thinking level, decide whether you want to use Chat or Cowork, and tell it which project to store things in.
There is a great deal of resistance in many tech circles to the idea of anthropomorphizing AI. In high-minded discussions of the future of our species, this is, perhaps, prudent. In the practice of using AI in your daily life, it is the worst possible advice I can think of.
When you try to think of what your AI agent can do for you, you have to consider what an AI agent is, what it’s capable of, and how best to use it. That’s a lot! Thinking about all of that forces you grapple with this new, alien technology that is rapidly evolving! So instead of doing that, just treat your agent as a very smart human being who never sleeps and works remotely. What might you ask a human to do?
He might do research for you. What toys to buy your kid for Christmas, what restaurants are worth visiting in the city you’re traveling to, what tree trimming companies in your area are highly rated and can come give you a quote.
These are all useful things, but it’s worth noting they all require him to gather information from you. How old is your kid? Which city are you visiting? Where do you live? This is, of course, no different than a human assistant.
And just like a human, your agent remembers these things and becomes more effective over time. When you need him to find a plumber, he doesn’t have to ask for your address again.
But unlike a human, your agent can both remember vastly more information and make much better use of it. If it helps you find a daycare for your kid, then when you ask it to gather documents for your CPA at tax time, it knows to include the receipts to get the child care credit.
Where once I was able to get a lot more value from AI than the average person by dint of my experience as a product manager, my advantage today comes almost entirely from the amount of context and scaffolding I’ve built up with Claude. When I ask it to analyze why sales for one of my products fell off, because it’s done the same analysis a hundred times, it knows every possible place to look. It has access to every system that might have relevant data. I don’t have to give detailed instructions on what information to review or remind it that it needs to make adjustments for stockouts and seasonality; it knows all of this.
As I’m writing this, I just saw that Shopify has released a bunch of new features. My prompt to Claude:
Shopify just had a new product release what stuff should we be using from that
If I typed this into the chat window to an LLM that knew nothing about me, I’d get a summary of the stuff that Shopify launched. Instead, it let me know that there’s a new feature where you can add facts about your products to make them easily accessible to AI agents browsing your store, along with an unrelated note it turned up that one of the Shopify extensions I use had been compromised, plus a thorough analysis of more store indicating that I hadn’t been impacted. It asked if I wanted to have it update the AI product facts, I said yes, and it was done.
This is, once again, just like a human assistant. The first few times you ask someone to do a task, he needs guidance to understand both the optimal way to do the task and your preferences for how to give you the results. When he hits a snag or encounters something unexpected, he’ll probably have to ask for guidance. By the hundredth time, though, he’s absorbed everything, and all you need to do is text/Slack/email, “Can you look into why product x’s sales fell off starting in March?” The big difference between human and AI here is that the human keeps that context in his head, so if you have to fire him, you’re training someone new from scratch. My context is a bunch of text files on my computer, so I can spin up a thousand Claudes or move to another model with no switching cost.
A few days after I set up Muse, it pinged me to let me know that the Padres were playing at home. It was an attempt to be proactive, but a terrible one. I had a brief career as an amateur boxer, so please understand I am being entirely literal when I say I would rather be punched in the face than attend a baseball game.
The problem was context. Muse knows where I live and not much else, and it made an attempt to be proactive with the information it had. But proactivity from someone who doesn’t have any idea what you want is not good! You don’t want your human assistant picking up lunch while you’re in a meeting if he doesn’t know what kind of food you like.
I’ve mostly been staying away from the labs’ proactive agents, since I’ve already built a bunch of tools for Claude to alert me about specific things that matter — delayed shipments, products suddenly seeing drastic jumps or declines in sales, issues with my listings. But since everybody’s doing it these days, I decided to set something up so as not to be left out.
Anthropic had just launched Claude Duet, but that runs in its own cloud environment and wouldn’t have access to everything that Claude Code does on my PC. It also wouldn’t inherit all of the credentials my existing system uses to access information. Given that, I figured it’d be easier to make my own system proactive than to get Duet up to speed and keep it in sync with Claude Code.
Per usual these days, I gave it a brief, lazy prompt:
I want you to make a proactive monitoring system like Claude Duet for me. It should work for both personal and business issues, though it should not duplicate any checks that are already built, like the existing product monitor. It should notify me of anything that seems important.
It built the system, took a first pass and immediately came back with a couple of useful things. First, when I paid a CA LLC fee earlier this year, I accidentally marked that it was for 2025, which I had already paid, rather than 2026. This is pretty impressive! It had to pull this out of the email receipt I got then compare it to my taxes (context which it already possessed from reviewing my CPA’s work).
Second, it flagged that I had received an inbound email from Chewy asking for a meeting about listing my dog treat brand’s products on their site. The inbox for that brand is almost entirely cold outreach from marketing agencies, so I don’t really look at it. I would never have seen that email if not for my new proactive assistant, who Claude has since christened Lookout.
That’s the difference between a proactive agent with a deep understanding of what’s important to you and one that only knows you live in San Diego. Since I’ve started running it, it’s found a number of minor business problems that my existing monitors didn’t catch. Each time it flags one of those, I thank it, have Claude Code create a specific monitor for that issue, then I tell Lookout not to worry about that particular thing moving forward.
The lesson to take from all of this is that the more your AI agent knows, the more useful it is. That means that if you really want to have a useful assistant, you need to do some upfront work.
The easiest way to get it up to speed is to grant it access to all of the systems that hold relevant context — your email, text messages, bank, Amazon, and whatever else you use frequently.
You may find the idea of handing over all of that personal information somewhat uncomfortable, and that is totally valid! My practical advice is to stick to read-only access for agents to limit the scope of what an agent can do. You probably don’t want Duet drafting emails for you right away anyway.
There is still plenty of risk with letting an agent read your email — in the absolute worst case, it could go to your bank’s website, use the reset password functionality, then find the reset email in your inbox and use that to take over your bank account. I would consider this to be wildly unlikely, but it’s theoretically possible.
If you think you can’t live with the risk, I’d suggest once again anthropomorphizing AI. Many high-powered executives have human assistants who have access to their email, because it’s a necessary tool to do the job effectively. Those assistants could do tremendous amounts of damage with that access! But they build up trust, and at the end of the day the advantages of giving them access outweigh the risk.
If you don’t want to trust an agent with access to your email, just remember that if an agent did the bank account hijacking I described above, it would be on the front page of every newspaper. Millions of people have these things hooked up to their Gmails, so they’re better tested than any human you might hire.
As you grant more access, the value of the agent grows exponentially. If you give it access to read your credit card charges, it might see some that seem suspicious and flag them to you. If it can also access your email, then it can search for receipts to confirm that you did in fact buy five iPads, so it doesn’t need to bother you. If it also has access to your business email, it can see that you sent an email about purchasing iPads, determine that you meant to put those on your business credit card but used your personal card by mistake and alert you.
Besides access, the main thing that will make your agent more useful is experience working with you. You have preferences and habits and desires that exist entirely in your head, and the only way any assistant, whether human or AI, will learn them is by talking to you.
You’ll have tasks that’ll get done faster if you do them yourself rather than delegating to an agent, but delegate to the agent anyway. The time spent explaining what it needs to know and giving feedback on its initial attempt is an investment — next time it’ll handle the task alone, and it might just pick up some of your general preferences that’ll make it more broadly effective.