Fun 32: A quick look at agents

Agents work while you don't

Chat waits for you to ask. An agent keeps working in the background, checking your email and calendar and coming back with things worth your attention. This session compares the new consumer agents (dots, Muse and Grok Bot), sets one up live by talking to it, and covers how to hand over work, email and data in stages.

Key takeaways

Agents are proactive where chat is reactive. Until now, AI did something amazing when asked and then stopped until you spoke again. An agent checks your email every couple of hours, flags what matters and prepares you for the meeting on your calendar. That takes real weight off your mind space, which a chat window never could.

You set up an agent by talking to it. Agents used to mean flowcharts of if-this-then-that rules. Now you describe the job in conversation, connect your calendar and email, and train the agent the way you would onboard an intern or a new hire. Each agent also gets its own computer, so it can browse sites, make bookings and do research while you sleep. Keep the number of agents under ten, because your brain melts well before that.

Build trust in stages, from crawl to walk to run. Start with an agent that reads your email and drafts replies into Gmail but never sends. Let it add shopping to a cart while you check out. Hand over more only as it proves it knows you, and decide how much data to give each company based on the value you get back.

Tools covered

dots: OpenAI's always-on agents with their own cloud computers that connect to apps and move work forward.
Muse: Meta's personal AI agent that helps pursue goals, take action across apps, and continue work in the background.
Grok Bot: xAI's always-on AI teammates that work across apps from their own computers and complete jobs end to end.
ChatGPT: OpenAI's AI assistant.

A quick look at agents, in full

A quick look at agents

A big transition in the AI space is the dawn of agents. The big difference between a ChatGPT kind of chat experience and an agent is that an agent can be proactive. It does things on your behalf in the background, the things you can do on your computer, while you sleep.

Agents work proactively

In the ChatGPT universe they just released dots, which is your personal agent. There are a couple of other good offerings. Facebook, believe it or not, released an incredible agent solution called Muse, and they're going after everyday consumers, people who already have Facebook or Instagram. Probably most people will end up with a Muse agent, because they inherit it from that relationship. I don't know if it's free, but I imagine it's very affordable.

The only problem is that Google, my former employer, and Facebook are two of the founders of surveillance capitalism. They make their money from your data, and I don't know that you want such a personal AI hosted by a company that harvests your data. They say they won't, but I don't know how much I would trust them. So you choose. It's not cheap, you're just paying another way.

Dots from ChatGPT is very, very good, and it's what I've been using every day for the last week or two. I think these agents are going to be a big part of how we interact with AI soon. It can feel abstract: what's the difference between ChatGPT and an agent? Until now, AI has been reactive. I want to do a thing, I tell it, and it goes off and does something amazing for me. As soon as it's done, it's done, and until I say something else it will not do anything. An agent can be proactive. It does things on your behalf in the background. If I look at my dot, I have four new messages. It's checking my email every couple of hours and telling me, "This came in, this is actually quite important, we've been talking about this." Or it checks my calendar and says, "There's a meeting coming up. Do you want to do this, or should I create something to prepare you for it?"

That proactive AI is really great. Parenting is a good example: there's so much mind space, because you constantly have to be aware of all these different things. AI up until now has never been able to help with your mind space. It can solve a problem for you, but I still have to hold the thing in my mind and get help with it. I can't say, "Could you just handle this part?" That's the big thing with agents.

You set up agents by talking

Agents have historically been really complicated to set up. Back in the day you had all these complex interfaces: if this happens and that happens, there's a noodle to this thing, then sign up to that thing. It's basically coding, but with a UI. Now agents are very simple to set up, and Grok Bot is maybe the simplest. Think of them as little coworkers.

Let's create a parent example. I'll create a new bot, and it greets me: "Hey Sebastian, good to meet you. I'm your new bot, ready to take something off your plate." I tell it we're creating an example bot. I'm not a parent, but we'll pretend. What I want help with is managing the day-to-day of my kids: sports practice, emails from the school, their calendar, maybe food planning. You're going to be my all-round assistant for the logistics around my kids.

This is how you set it up, with a conversation. You just talk to it. My calendar is already connected, so I can start there. You can connect all the apps you use day to day, like Gmail and Drive, or Notion and Slack. For a parent, it's probably mainly email and calendar. To make it real, it asks me to tell it about the kids: names, ages, whatever. Then I'd ask it to explore how it helps me throughout the week, and which touch points are useful.

That's the cool thing about where agents are today. It's just a conversation, and you can think of it as an ongoing relationship, the same way you would with an intern, an assistant or a new hire. How would you get them up to speed? How would you train them to be most helpful to you?

My brain starts to melt at around ten agents, so keeping it under ten is probably a good idea.

Each agent gets a computer

One thing to know about Grok Bot and a lot of these agents is that they have their own computer. That's a bit abstract, but it means they can do things you can do on your computer, like while you sleep. It can go to websites, make bookings, do research and use specific software, which gives it a lot of power.

Here it gives me a packed lunch plan and a shopping list. That isn't super magical, and you could do it with ChatGPT. But because the agent has its own computer, say you do online shopping. In the UK we have a store called Ocado. The agent could log into your account, add the items to the cart and take you to checkout, and then you pay. You could probably have it pay too, if your card details are saved in the account.

The way to start is something we call crawl, walk, run. Start small: it adds things to your cart and you do the checkout. As you build trust, you can start to walk, and then run. Don't go straight to running. Muse is built more for consumers, so it can take your credit card and run with it, but you really have to give it permission and you can set the guardrails. What they call agentic commerce is definitely going to be a big thing. It's predicted to be a huge part of purchases made on the internet in the near future: I need the thing, go buy it.

What you see here isn't magic in itself. The magic is that it can happen proactively. It knows it's Tuesday at 5pm and we need to figure out shopping, so it comes to you with a notification on your phone and you handle that one thing.

Agents run work and email

Take a bot that just watches your email for one purpose: somebody sends an inquiry, a booking request or a complaint. It can stay on top of that and draft replies from your templates for that kind of message. Or say you want to stay up to date on something specific without spending all your time searching Instagram and Twitter. It's easy to create a bot that keeps you informed. This one tracks new research and news in the AI space, which is very useful.

Here too it's crawl, walk, run. Email carries a lot of signal about your life, so it's really helpful for the agent to have that context. I don't want it sending email, though it can, but I love that it can draft an email and put it into my Gmail as a draft. My rules for my agents are: read my email as you need, whenever you need. If I ask for a draft, either show it in the chat so I can see it, or send the draft to my Gmail so I can send it. Never send an email for me.

As my trust has increased, there was a first time when I said, "That draft looks good, can you just send it?" Now I do that more and more, but I never send an email without reviewing the draft and making sure it sounds like me and says exactly what I want. Be careful and build the trust. Don't trust anything at the start. As you build the relationship and it knows you and understands what you need, you can give it more responsibility. Give it access to your email only if you're comfortable with it. I wouldn't give Muse my email, but I would give ChatGPT my email. It can be very useful.

Try agents, keep data safe

That's a good question. Almost all of the platforms let you turn off training on your data, and I recommend going into the settings and making sure it isn't training on your data. For general health-related or highly private things, when a platform has an incognito feature, go for it. It's also an appetite for risk. It has access to everything about me already, and as long as it provides a lot of utility, I'm pretty lax about it and give it access to whatever it needs. Financial data is highly sensitive, but if it adds value by helping me understand my finances, investments and savings, it's worth it, because otherwise I just don't have that kind of oversight.

With all of these, you're working out which companies you're willing to give your information to, how much of it, and whether it provides actual value. The trade-off with Facebook back in the day was unclear: was it providing a lot of value, and how much data was it taking? This feels like a more equitable trade, but we want to make sure it's worth it.

If you want to start playing with agents and you don't have access to a dot yet through your ChatGPT subscription, I'd suggest Grok Bot. See whether it adds value, for kids' stuff for example, which is a perfect use case, and then build from there to your business and your day-to-day. It's an exciting new paradigm.

The one thing to be careful about is having too many bots. Your brain can only handle so many people, the same way it would melt if you had to deal with forty different people at work doing different things. With ChatGPT we only get the one. It can build other agents or use other threads to do other tasks, but you talk to one assistant and it orchestrates across different chats. That's quite nice: you talk to one, and it figures out who it needs to talk to.

Really, just try playing around with it. It's a new way of interacting with AI, and quite an enjoyable one. AI becoming more proactive means it can be more useful in our lives and, hopefully, offset some of the mind space that our day-to-day lives require.