I did not expect ChatGPT’s Dot and Meta’s Muse to be so good at normal things.
Not complex demonstrations built to show what an AI agent might eventually do. I mean the ordinary tasks that take time and attention throughout the week: booking reservations, checking flights, responding to emails, sending text messages, monitoring important messages, and remembering things I did not want to lose track of.
I gave both agents similar work. In most cases, they handled it with very little guidance from me.
That was the first surprise.
OpenAI describes Dot → as an always-on agent inside ChatGPT. It is powered by GPT-6 Astra, has its own cloud computer, works across applications a user chooses to connect, and turns to the user for decisions that require judgment.
Meta describes Muse → as a personal agent that runs in its own secure cloud computer, called Muse Secure VM. It can use a browser and connected applications, continue working after the app is closed, and request approval before sensitive actions.
Those descriptions explain how the products work. Using them showed me what those capabilities feel like in practice.
I gave them real work
I did not spend much time creating elaborate instructions. I told them what I needed in roughly the same way I would explain it to another person.
Find a reservation. Check flight options. Reply to this email. Send this text. Let me know when an important message arrives. Remember this for me.
During my testing, both agents were able to work through most of the intermediate steps without requiring me to direct every action. They navigated websites, used connected applications, entered information, and stopped when they reached something that required my involvement.
In the tasks I ran, entering payment details was one of the points where I needed to step in. Both products also support approved purchases through certain saved-payment or wallet flows.
The value was not that either agent could perform one extraordinary task. It was that they could complete a series of ordinary ones with very little management from me.
My mom’s birthday dinner made it feel real
I was planning a dinner reservation for my mom’s 75th birthday weekend. I told Dot what I was looking for, and it helped find a restaurant and book the reservation.
The work continued after the booking was confirmed. Dot contacted the restaurant about the birthday celebration and handled the follow-up conversation about special arrangements for her big day. Having a reservation was one part of the task. Making sure the restaurant understood the occasion and working through the details mattered just as much.
That sequence is what made the experience feel different. Dot carried the context from the initial request into the conversation that followed. I could explain what I wanted for my mom, then have it help coordinate the details across the booking and email exchanges.
In the regular chat workflow I had been using, I would still have needed to evaluate the options, visit the booking site, complete the reservation, find the restaurant’s contact information, send the message, and track the response myself. Dot worked across the entire process for me.
Dot started with more context
Dot had one clear advantage in my testing: it began with more context about me.
OpenAI says → Dot can receive saved memories and recent conversation context from ChatGPT, with sharing continuing after setup when Memory is enabled. It can also build its own context from conversations and connected applications. That does not mean it has unrestricted access to every conversation I have ever had.
What I experienced was an agent that already recognized more of my established preferences, interests, likes, dislikes, and ways of working.
I did not always have to explain why I might prefer one option over another or repeat information ChatGPT already remembered.
Muse also has persistent memory. Meta describes memory files → that users can inspect and edit directly. My relationship with Muse is newer, so it had less history with me to draw on during my testing.
This does not make Muse less capable. It means Dot had a head start.
In an agent, memory is more than personalization. It reduces the amount of briefing required before work can begin.
Access can matter more than intelligence
Muse gave me an easier path to my calendar during my original testing.
At that point, Dot was connected to my Google Calendar account, but I could not connect it to the iCloud calendar I use through Apple Calendar on my MacBook and iPhone. Muse connected to that calendar without a problem. That describes the connections available in my setup at the time.
Since then, I’ve found another route for Dot to reach my Mac apps. In ChatGPT desktop version 26.1007.41230, local Mac access is now available in my setup. That is new for me, and it changes how I think about the comparison. I’m describing when I noticed it, rather than claiming this particular build introduced the feature.
OpenAI confirms → that local access is optional and starts turned off. You enable it in the desktop app on the computer Dot will use. That computer must remain online with ChatGPT open, and the work is subject to the permissions and tools available there. Going offline makes the computer unavailable; it does not revoke the access you granted.
You can also request work on that connected Mac from the ChatGPT mobile app when Dots is available on your account. The task still runs on the Mac. Using your phone to make the request does not remove the Mac connection requirement. Mobile web is not supported for Dots.
This local route also differs from connecting a cloud service. Dot has its own cloud computer for work that does not require my Mac. I haven’t established that local access gives me the same calendar workflow I originally used with Muse, but it gives me another option to explore.
The two agents performed similarly across many of the tasks I tested. The calendar difference had more to do with practical access than intelligence, and that access is evolving.
People will experience an agent through what it can actually do in their lives, including when and where it can reach their apps.
Their approval models felt similar
Both agents performed work through dedicated cloud computers, although the companies implement and describe those environments differently.
Their approval experiences felt similar to me.
They could complete much of a task independently, but they paused when an action required my involvement. That was especially clear when a task reached payment information.
OpenAI gives Dot users controls that determine whether it can act without asking, act when pre-approved, ask before acting, or hand an action back to the user. Those custom rules work within built-in safeguards that cannot be overridden. Meta says Muse requests permission before sensitive actions such as sending an email or making a purchase, with a separate Sentinel system enforcing its permissions.
OpenAI documents approved purchases → using cards saved on merchant websites. Meta documents → purchase approval for saved payment methods and Stripe Link wallet payments, which use single-use cards. My experience reaching a payment handoff should be read in that context.
In my use, both approaches gave the agents enough independence to be useful while preserving clear points where control returned to me.
I did not need them to act without limits. I needed the limits to be understandable and predictable.
That predictability is an important part of trust.
Muse feels easier to call on
The biggest experience difference may be the simplest one: Muse is its own application.
I prefer that.
When I open Muse, I am opening an agent because I want something done. The purpose of the experience is immediately clear.
Dot benefits from being integrated into ChatGPT, but it is also one capability inside a much larger application. I already use ChatGPT for research, writing, questions, projects, and many other forms of work. Reaching Dot means entering that broader environment first.
ChatGPT is a substantial application because it does many things. Muse feels lighter because it has a more focused job.
One is a place where I regularly think with AI. The other feels like a place where I delegate work to AI.
That may sound like a minor interface distinction. In practice, it changes how quickly I remember to use the agent and how focused the interaction feels.
I would like Dot even more if it had its own dedicated application or a more distinct surface separate from the rest of ChatGPT.
The real test is ordinary usefulness
I cannot say that one agent is definitively better.
Dot had a head start in accumulated context. Muse gave me an easier calendar connection in my original setup and, for me, a more focused application experience. Dot’s connected Mac access adds another option, with its usefulness depending on the apps, permissions, and connection available. Their ability to complete the everyday tasks I tested was surprisingly similar.
These products are new, and their integrations and capabilities will continue to change. This is a snapshot of my experience using them now.
The more important discovery is that neither one required sophisticated prompting or constant direction.
I could describe what I needed, let the agent work, and step back in when a genuine decision was required.
That is where personal agents are becoming useful. Not in dramatic demonstrations, but in the small tasks that quietly consume time and attention.
For me, both products have crossed an important line. I no longer only ask them questions.
I hand them work.
Did you like this?



