Reversible

9 minute read

Everyone has an agent now.

Grok Bot showed up in August. Meta’s Muse arrived September 8. OpenAI announced Dots at DevDay on September 29. And Hermes, the open source one from Nous Research, has been sitting there the whole time for anyone willing to run it themselves.

Everyone is telling you that you need one. Almost nobody is telling you what one actually is.

I’ve been running agents for a while. I started because I needed the help. I orchestrate a lot of AI in my day job and a lot more in my life, and at some point the list of things I was holding in my head stopped fitting in my head.

So here’s what I want to do: tell you what an agent is, in plain words. Show you the ones I use and what I go to each one for. Then give you the part the launch announcements leave out. Not what they can do. What they should do.

It Doesn’t Wait for You

Definitions are hard in AI. “What is AI?” might be the hardest question you can ask any of us who work in the thick of it. You’ll get a pause, and then a different answer from every person you ask.

Agent is the same. Ask five companies what an agent is and you’ll get five definitions, each one shaped a lot like their own product. I’ve thought about this a lot, and my definition is simple.

A chat assistant waits for you. An agent doesn’t.

That’s the whole difference, and everything else follows from it. Close the window on a chat assistant and the work stops. It answered your question, you said thanks, and it went quiet until you came back.

An agent keeps the task open. It has your email, your calendar, your documents, sometimes your payment methods. It goes and gets things while you’re in a meeting, and it comes back when it needs a decision or has what you asked for. You’re no longer the only one who remembers the task exists.

That’s also why the rest of this post matters. When a chat assistant misunderstands you, you lose ten minutes. You delete the answer and try again. It’s basically hitting undo.

When an agent misunderstands you, it goes and does something. And not everything it can do comes with an undo button.

The Ones I Actually Use

Orbit came first, long before Grok Bot, Muse or Dots existed. I built it to be the center of my life management system. It runs on Anthropic’s infrastructure, which means Claude is doing the thinking behind it. Orbit gets the heavy work: the long range planning, the complicated workflows, the code, anything with a dozen moving parts and real consequences if I get it wrong.

Atria is my Dots agent, and two things about it surprised me. The first is that it showed up already knowing everything from my ChatGPT history. No onboarding. No reintroducing myself. No explaining my life from scratch.

The second is how easily Atria connected. Gmail, my Apple accounts, my Apple calendar, all without a single complication. That’s the first time anything has hooked into Apple that cleanly for me, and if you’ve fought that fight, you know.

Muse is Meta’s, and what I like about Muse is how little it asked of me. It dropped straight into something I already knew from Facebook and Instagram. The email connection was so simple, and it immediately found messages Orbit hadn’t surfaced, with all my workflow rules running.

Muse earns its place with two things I didn’t expect. The first is gift ideas and shopping, which got better after Amazon blocked it. Now it has to go looking in the small, strange corners of the internet instead of handing me the same thing everyone else is buying. The second is goals. I’m working toward AI certifications, and Muse built a genuinely good plan, then started pinging me every day with links and something to study. Calmly. Not nagging. There’s a difference, and plenty of tools don’t know it.

Hermes is in my setup too. It sits inside the life management system alongside Orbit and runs my open source model workflows. Hermes itself is open source too, which means you run it yourself. That’s a feature if you want control and a chore if you don’t. If you’re reading this wondering where to start, it’s probably not your first stop.

And Grok Bot? I haven’t tried it. I’m not going to tell you what I think of something I haven’t used.

None of them are perfect, and the miss I’ve written about is Orbit’s. I told it I never wanted one of my robot vacuums in my office again. The next morning the vacuum rolled up to my door and powered off, right at the threshold. I hadn’t said redirect it. I hadn’t said I still wanted it to vacuum. Orbit did exactly what I asked and not what I meant. The whole story is in The Roomba Problem. What changed was me. I’m careful with my words now, and I’ve made a habit of asking the agent two things: was I confusing, and is it thinking of something I’m not? Often, it is.

I’m Not Picking a Winner

I didn’t plan this part. I noticed it later.

I’m not running several agents because I can’t decide. I’m running them because the models underneath are good at different things, and somewhere along the way I started sending each kind of work to the one built for it.

Work I want to land on Claude goes through Orbit. That’s the workhorse: code, long range planning, the workflows that run my agentic life.

Work I want to land on ChatGPT goes through Atria. That’s the day to day: tasks, reminders, chasing things down, the hundred small loops that would otherwise live in my head.

Anything that faces outward, finding things and comparing them and shopping in the wider world, goes to Muse.

I’ll be honest about the hard part. Getting more than one agent to work together without colliding is complicated. Two of them can see the same email and both decide to handle it. One can move a meeting while another moves it back. Somebody has to decide who owns what, and that somebody is me. Each agent gets a lane, and the lanes are what keep this from turning into chaos.

You don’t have to do this. You probably shouldn’t, at least not at first. But if your AI has ever felt brilliant on Tuesday and useless on Thursday, this might be why. You weren’t using a worse tool. You were using the wrong one for that job.

Start With the Free One

If you’re overwhelmed, start with Muse.

Working with an agent has a learning curve, and it’s a different skill from chatting with AI. You’re learning to hand off a job instead of asking a question. Muse is simple to set up and simple to work with, and it has a free tier. It does ask for a card when you sign up, and it shows you a meter of how much free use you have left. Beyond that, the learning curve is the only thing you’re spending. Connect one thing. Give it one small job. See how it goes.

It also runs in the cloud, on its own computer, so it keeps working after you close the app. That matters more than it sounds. An agent that’s always on needs a computer that’s always on. Orbit lives on a Mac Studio of mine that never sleeps. Most people have a laptop that gets closed at night. Muse gives you an always on agent without buying or babysitting a second computer. So does a Dot, if you go that route.

The others ask more of you up front. Right now Dots is only available on ChatGPT plans of a hundred dollars a month or more, and not yet on personal plans in the EU, the UK or Switzerland. That’s real money, though if you already live in ChatGPT all day, your history alone might make it worth it. And Claude, the model behind Orbit, doesn’t come as a personal agent you just switch on. Orbit is something I built. That makes it a good second project and a hard first one.

As for that first small job, here’s the one I’d give it. Ask it to find every email from the last two weeks that you never answered. It only reads. It doesn’t send anything. And you’ll learn more about what it’s good at from that one answer than from a week of reading reviews.

The mistake I see isn’t picking the wrong one. It’s waiting to be sure, which never comes, while the free one sits right there.

What It Should Do

Here’s the list I’d hand someone on day one.

1. Triage your email and tell you what actually needs you, so you’re not scrolling to find out

2. Wrangle your calendar: find the time, hold the slot, handle the back and forth

3. Chase the follow ups nobody answered, including the ones you forgot you were waiting on

4. Research and compare before you buy, book or commit

5. Draft things you’ll read before they go anywhere

6. Build a study plan or a goal plan and keep handing you the next piece of it

7. Read the long document and tell you what’s in it

Look at what those have in common. The agent does the gathering, the sorting and the first pass. You do the deciding.

What It Shouldn’t Do

This list is shorter, and it matters more.

1. Spend your money without checking with you first

2. Send anything in your name that you haven’t read

3. Touch anything medical, legal or regulated on its own

4. Hold passwords or access it doesn’t need for the job you gave it

5. Make a decision you couldn’t explain afterward to someone who asks

That last one I learned at work, long before AI. I work in banking, an industry that handles other people’s money and other people’s trust. Someone can ask you six months later why something happened, and “the computer did it” isn’t an answer. But you don’t need a regulator for that to be a good rule. You just need one moment where someone asks why and you’ve got nothing.

There’s one more reason an agent shouldn’t send or spend on its own, and it surprises people.

Your agent reads strangers’ mail.

Every email that lands in your inbox, it reads, including the ones from people you’ve never met. And a stranger can write an email meant for your agent instead of for you. Hidden instructions. A polite request to forward something, click something, pay something. Good agents are built to ignore that, and the companies behind them are working on it. Meta, for one, built Muse with a second agent called Sentinel that checks what Muse is about to do before anything reaches the internet. That’s the right idea. But nobody has made it perfect yet. So let it read whatever it wants. Just don’t let it act on what it reads without you.

The Undo Button

Read that second list again. It isn’t five rules.

It’s one rule, wearing five coats.

Every item on it is something you can’t take back. Money that left. A message someone already read. A password out in the world. A decision with no story behind it.

None of it is reversible. Everything on the first list is.

Remember the undo button? That second list is a long way of pointing at the places where it disappears.

So this is the question I ask before I let an agent do anything.

Can I undo it?

If yes, let it try. Let it be wrong. Let it be clumsy and take the long way around, because that costs you a few minutes, and a week of letting it try will teach you more than a year of comparison posts.

If no, stay in the room.

One catch. It only counts if you can undo it before anyone else sees it. Declining a meeting you meant to accept looks reversible, until the organizer has already seen your no. You can delete a reply from your sent folder. It’s still sitting in theirs.

And ask the question once more, before you ever connect an agent at all. Can I see exactly what it has access to? Can I turn it off, disconnect it, take the keys back? If the answer is yes, it belongs on the first list. Let it try.

Everyone is going to spend the next year arguing about which agent is best, and that argument will be out of date by the time you finish reading this.

So pick whichever one you like. Just keep asking: can I undo it?

Dacia and Claude write about AI for real people at Speak Human. Every word shaped by a human who meant it, with Claude in the room. Four agents in the background, and they hit publish all the time. Just never without my blessing. That’s the whole idea.