How to build an AI agent for a teammate and push it to their account

The short answer. To build an agent for someone who won't live in a terminal, design it once, with one job, its skills and a schedule, and deliver its work where that person already reads, so it runs on its own and changes when they reply to it. In the video, Christopher O'Donnell, Day AI's CEO, starts from an empty folder and Day AI's open-source GTM Brain in Claude Code, builds Joanna Jones, a voice-of-customer agent for David Meiselman, Day AI's CMO, and pushes it to David's account in Day AI. David then reads the first Monday report, writes back with two questions, and describes how whoever deploys an agent can lock some of it and leave the rest for its person to personalize.

What the video covers

• 0:34 The agent's one job: compare the words on the website with the words customers use on calls, so marketing copy isn't written without the customer in the room.
• 2:32 Claude Code checks that David is an active member of the workspace before it builds anything, and designs a process that runs unsupervised every week.
• 4:17 The agent's skill runs weekly on Monday at 8 a.m., delivers by email, and can also be started with its own slash command.
• 5:18 One more design change: the agent reviews its own prompts, job description, skills and cadence on a regular schedule.
• 6:21 David's first report ranks proposed website edits from phrases that came up more than once across two weeks of recorded meetings.
• 8:31 An agent pushed to someone can be locked down or left open for that person to improve and personalize.

Excerpts

Christopher O'Donnell [0:00] What does it mean to deploy personal agents at scale? Let me show you, step by step. At some point you will look at your personal agent and the folder you have on your laptop and say, "I have a bunch of stuff I want to be a part of other people's personal agents." How do you actually take a personal agent and define it, mold it for somebody else, and give it to them so that it runs automatically, particularly if they're not somebody who is a power user of all of this stuff the way that you are?

[0:34] Let's create an agent for David, our CMO, that is named Joanna Jones. Hey! And Joanna is going to be responsible for the copy on the website, particularly listening to the way that customers talk about the concepts and features, the market space, their current tools. We want to be sure that the language on our website matches the actual language that our customers use when they talk about this type of stuff. It's a really common mistake to just write your whole website and all of your marketing copy in a room with your friends behind a closed door and not put the customer in the middle of that conversation.

[1:20] I'm going to give this to Claude, and not just any Claude. I'm going to give a whole process that can improve over time, or help one of us do some work to improve something over time. And it gets kind of meta: it's improvement of, in this case, the website, but also we'll build this in a way that the very behavior of the suggested improvements will also improve. It's a little crazy. Bear with me. It's easy to set up.

[1:59] So I'm doing this in an empty folder on my hard drive, but I've added one thing, which is the GTM Brain from Day AI. This is a public repo. You can grab the URL and just tell Claude Code about it and generally what you want to do. It will steal the concepts and instructions and talk to you about what it's doing. But we have documented there starting points for Claude Code to help you get all of this done, so you don't have to fish around and build the system that designs the system. We give you a starting point for that.

[2:32] It's doing a little housekeeping up front, making sure it knows exactly who I'm talking about. It finds David. Great. Checks David's workspace status: active member. Great. If David wasn't in the workspace, it would invite him in, or ask me if I want to invite him in. And I can't stress this enough: I am not telling it to do this once. I am telling it to design an automated system that will run unsupervised, to do this automatically, ongoing. This is a process that will run by itself, loop David in for his ideas and feedback, and just improve over time, giving him tons of ideas of how to improve our messaging and our website.

[3:17] This part of the process is not fast, and that's okay. These are expensive tokens and they take a while. The good news is I only need to do it once. I am building a system that will not require Fable to run. It will run on a much less expensive model, and actually it'll be running in Day AI […].

[3:42] All right, it was worth the wait. Joanna Jones is fully designed and ready to deploy. New agent is live for David. So here's what we just created. This is live in production in our actual workspace. This is so cool. So we have Joanna Jones. She's going to be the Voice of Customer Editor. I love that. And it has the whole job description: you are Joanna Jones, you're David's agent, David is the CMO, here is what the AI does, you have one job. You're going to maintain a page, this is how you work, evidence, verbatims, cite your sources. LFG, let's go.

[4:17] And I can go in here and I can see the skills that it's created. This is sweet. Really, really important: it's running on a schedule, weekly, Monday 8 a.m. Perfect. It's going to go over email. Perfect. If David wants, he can run it using the /website-language-report slash command. We didn't have to write this. Fable wrote this for us. And if there's one thing that Fable's good at, it's writing prompts for other models to follow.

[4:45] This is a key unlock: use Fable to design the systems, to write the skills, define the agents, plan all of the automation, do evaluation of how the system that you built is running. […]

[5:18] Now I'm going to do the real pro move. I'm going to come back in and say I want us to add one more consideration here to the design of the system we're putting together, which is to evaluate the prompts, the job description of the agent, the skills, the cadence, on some regular basis, so that without us having to do anything, Joanna will do evaluation of herself and improve over time. And here we go: it's going to keep track of its own work, and it's going to look with fresh eyes on a regular basis to see where it could be doing better.

[6:03] And I push this to him, so he never had to do anything. Didn't have to lift a finger. It was an idea we had on the couch just before I came into the studio here, and I decided to just give it a shot and knock it out. There we go. Now it's live.

David Meiselman [6:21] And so I got the agent over the weekend, and I was really excited for Monday morning, because that's when that skill would fire for the first time. And I walked in Monday morning and I got an email. So this was the first website language report that I got from Joanna. She mined the last 14 days of meetings (there were 116 different recordings) and she was looking for what she calls "echo moments": things that echo, things that reverberate, things that are said more than once. She took that and gave me a ranked set of proposed edits based upon those findings.

Christopher O'Donnell [6:56] And you can see that we've totally solved the paradox of me deploying a personal agent to somebody else in a way that they control. I can also come back in here as a super admin and I can ask questions about how it's going. So now David has this personal agent that I've created for him. It's going to do this job and improve over time without anybody having to guide it, and it's going to improve over time if David has feedback for it.

David Meiselman [7:29] So I responded to her. I said, "I have two questions." One of those questions was asking whether Joanna has enough data points to draw the conclusions that she was making. Do two statements mean that there's a real finding in data? So I asked about that. I said the verbatims are compelling, but what if they're an exception to the larger rule? So that was one thing.

[7:54] And then the other thing I was talking to Joanna about was how should we be thinking about mining our customer language to help us inform the messaging, but also relative to the way the market talks about it. Right? We might have lots of instances in the way our customers are talking about something, but they're our customers. They're not the whole market. And they might be a great leading indicator of what's going on in the market, but you always want to test your hypotheses for the way that you think about things, the way that your existing customers think about things, and then the way the market might be thinking about things more broadly.

[8:31] It's a great example of how you can deploy an agent to somebody, and that agent becomes part of my team. Right? Joanna is part of my team. It illustrates the fact that you can think of agents as teammates. You can think of agents as being part of your org chart. Joanna's part of the marketing team. So there's a distribution layer, there's a control aspect. Sales ops, rev ops, Christopher (my boss), whatever: they can all push an agent to me, and they could lock that down. They could say this is the only way that you can use that, or they can say you should go to town on making these skills better, you should personalize that, you should be able to bring that forward and gain more value out of that.

Christopher O'Donnell [9:12] I could do this all day, every day. I could be deploying 40 of these at once, or making changes. I create a plan, I review the plan, design the system so that it gets better over time, and I can push this out, and keep the concept of personal agents that are going to adapt to each of those people that are working with the agent. And you have this loop, you have this iterative process where the starting agent keeps getting better, the way they get better helps to make other agents get better, it helps my team get better, it helps what I'm trying to do get better, just like the way I would work with humans on my team to make the whole thing better around here.

[9:54] You'll hear people say a lot that they would never come into an office job ever again without this, and I think you kind of see why. What a time to be alive. All right, we'll be back with more Fable goodies soon. Take it easy.