Recently a consulting firm sent me an invoice. My agent saw it, reviewed the attached PDF, found an email conversation that listed a different number of hours, highlighted the discrepancy for me, and offered to draft and send a message to the firm questioning the bill. I didn’t ask it to do any of this. Honestly, I likely would have missed the discrepancy myself.

That’s what agents are supposed to feel like. With the launch of Meta’s Muse and OpenAI’s Dots, talk of the year of agents has really gone into overdrive. However, most people are using their agents like the chatbots that came before them (namely a nicer interface to something like Google search). That’s not surprising: when OpenAI studied how people use ChatGPT, nearly half of all messages were people asking for information, more than were asking it to actually do something. We’re bringing chatbot habits to a very different tool.

Agents really are a different experience - if you understand them and use them properly. The best version is less like a search box and more like a chief of staff.

What makes agents different

Agents differ from traditional chatbots in two key ways.

First, they can take actions on your behalf. They can not only research the best restaurant options for dinner with your friends, they can make the reservation too. Some of these capabilities have also been included in the chatbot versions, but this is a real shift in capabilities from even a few months ago.

Second, agents can be proactive - highlighting information for you without being asked, like my agent catching that invoice.

How to get the most from agents

Since agents can do different things from our chatbots, using them effectively means learning to use them differently too. Here are my top three pieces of advice for becoming an agent power user.

First, think async. It’s really wonderful that most agents interact through traditional messaging. Just like with a human assistant, you don’t want to sit around and watch your agent work. Just send the request and then keep doing other things until you get a response. In the meantime, feel free to send a second (and third!) request.

Second, have it build something for you. My favorite example of this comes from JJ Englert, who taught his Muse his TV and movie preferences and had it create a living document recommending movies to watch. It even checks new releases in theaters and on streaming services to recommend something for him - and if he finds something he wants to watch it can book the tickets for him directly. No need to find the app for that on the app store - just ask your agent to do it!

Third, give it a schedule. I have an agent that looks at listener metrics for my What the AI?! podcast every week and gives me a report Monday morning. I can review, ask follow up questions, and even give it some work to do based on the findings. I used to set aside time to compile and review that data myself - now my agent does that and I can move on to thinking about what to do based on the data.

What we need next

Agents so far are useful, but we are only scratching the surface of what’s possible. The agents that ultimately win usage and loyalty will need to address several challenges for users adopting agents today.

First is trust. I’m willing to go out on a limb and give (some of) my sensitive personal data to multiple agents. But not everyone feels that way, and the provider that can effectively earn your trust to share payment details and email sending permissions will have a huge advantage here.

Second is shared context and content across devices. An agent on your phone doesn’t have access to files or integrations on your laptop. This problem only gets worse when you’re using one agent on your phone and another AI tool on your laptop (for me this is Muse and Claude).

Today’s power users are running mini data centers to make up for this. I have a headless Mac mini right now with ChatGPT, Claude, and Muse always running so they can access my files synced via iCloud and an open source version of gbrain (https://github.com/garrytan/gbrain) to provide shared context across devices and agents. That is WAY too complex. Making this easier will make agents far more useful for most users.

The best version of an agent is something like a chief of staff. You can ask it anything, and it has the right context to take the right actions on your behalf. It can kick off whole new projects, or watch the projects unfold and bring you the questions that need answers to keep things moving.

There is so much potential here, but we are just in the early stages. Today, an agent catching a bad invoice feels like magic. Before long, it’ll just feel like having a great chief of staff, and we’ll wonder how we ever worked without one.