How to build an agent loop that finds the deals a CEO can help move

The short answer. Give the model the whole deal first, then ask it for a recurring system that does the work every time the situation comes up. In the video, ChatGPT and Claude with only Gmail can't tell why a deal is stuck, and Claude connected to Day AI over MCP finds the deal's meetings, opportunity details and Slack threads. Christopher O'Donnell then starts Claude Code from GTM Brain, Day AI's open-source Claude Code harness, and has it design and deploy a weekday loop: an agent sweeps every open deal for moments where the CEO can help, ranks them, drafts the move, and reviews its own results each week.

What the video covers

• 2:36 Claude with only Gmail connected searches the inbox for a long time and drafts from the last email, because the context for a company-level deal sits in meetings and Slack it can't reach.
• 4:08 With Day AI connected, Claude finds the meeting recordings, opportunity details and Slack conversations, each with the sharing rules people chose, and gets much closer, though the draft still misreads why the deal is stuck.
• 10:16 In Claude Code, with GTM Brain loaded, the request changes from one email to an automated, recurring system that keeps state in Day AI pages and improves with and without his feedback.
• 15:00 Claude Code records the work as an initiative and waits for approval before deploying, the same plan, approve and verify steps engineering teams use with AI.
• 16:02 The deployed loop runs every weekday at 7:15: an agent checks open deals across three pipelines for six moments, including deals gone quiet, slipped close dates, named blockers and closed deals that warrant a note from the founder, then judges which ones need the CEO, ranks them and drafts the email.
• 19:08 On Fridays the loop reviews the quality of its own work and updates its instructions, and Claude Code gets a command for checking how the loop grades itself.

Excerpts

[0:00] Claude Fable 5 truly has changed sales forever, and I'm going to show you exactly how and where you get started. Everybody is talking about this Mythos class of models from Anthropic. […] What does it mean for us in our daily lives, and what does it mean for salespeople? I want to show you exactly how to use this.

[0:26] If you've read anything online about how people are using Fable and how the creators of it use it, you'll notice that you're supposed to use it pretty differently from every other model you've ever used in ChatGPT or Claude. It's really intimidating at first and extremely simple once you get going with it and you make the shift in how you think about it. The funny part in all of this is you don't have to use Fable very much at all. A little bit goes a long way, and most of the time you can still use these fast, inexpensive models that we're used to. I want to start at the beginning, the way you first used it on your first day using ChatGPT, all the way through the kind of black belt level of applying AI to a sales job, and then show you the future, which is available to us now using Fable, and it changes everything. Let's dive in.

[1:11] All right, I'm going to use a real example. I have a deal that's a little hung up. I think it's in good shape, but I probably could do something to help. I heard about it around the office. I want to write the CEO an email. And we're going to start where a lot of people start. Maybe this is where you start. I'm going to open up ChatGPT and just dive in and say, "Let's write a nice email to […] to unstick the deal." And let's see how far it gets.

[1:37] It's asking me what's blocking the deal. That's the first clue that we're still at kind of white belt level. It should know what's blocking the deal. But let's play along, say […]. What outcome do I want? Resolve concerns. That sounds good. What tone should the email use? Okay, so it doesn't know what tone I use when I write an email from founder to founder. That seems pretty white belt to me. That's fine. Friendly and casual. This is not actually an email draft. I can open this in email, but it doesn't have an actual draft. It doesn't have information about who […] is. It doesn't know what company he's part of. This is totally generic. I'm going to have to basically copy and paste this. I'm going to have to look up […]. It doesn't say anything about the actual issue at hand. It doesn't know anything about my team or his team or the nature of the relationship. This is very, very basic. This is too basic and actually not worth my time.

[2:33] Let's try hopping over to Claude and doing this with some of the right data sources connected. So in Claude, I have a couple of things going on. In my connectors, I have Gmail and I have Day AI as well. We'll see what it wants to use. All right. I have Gmail turned on. I don't have Day AI turned on. It'll be interesting to see how Claude handles this. "Let's write […] a nice email to unstick the deal." We're using Opus 4.8. We'll see how it does. It doesn't have any CRM or customer memory details. It's saying, "Let me find the thread." It's using Gmail, but Gmail only. The Slacks, the meeting recordings, everything: it's not available. It's kind of hunting around in the dark here.

[3:07] This is still actually very much white belt level. It has some data access. It could theoretically write an email. Maybe we give it credit for being yellow belt, but this is not it. When it searches Gmail to try to figure out who I'm talking about, you can see this is taking sort of an awkwardly long period of time. Theoretically, since Claude is connected to Gmail, it can hunt around and think and spend the tokens and kind of waste my time trying to figure out who […] is, who […] is, what's going on, what the deal is. It takes a long time. And there's no guarantee at all that it's even going to come up with the right answer.

[3:50] And in this case, I know it's not going to come up with the right answer, because this is not the continuation of a conversation in email that I had with the person in question. This is something that's happening at a company level, which means that Claude just doesn't have a chance here at all. So it comes back and it notes it doesn't have access to any of the right data. It's working off of my last email to him and nothing else. And so the email draft that I have here has nothing to do with what I originally wanted to do in the first place.

[4:21] Now, let's try this in Claude, but with the right data connected. If I turn on Day AI, and I don't need Gmail at this point, and I ask the same thing, and I say, "Let's write […] a nice email to unstick the deal." Now, let's see what happens. It's doing that same planning step, considering how it might go about searching to find the right context, how it's going to approach writing this email, and then pretty soon here, I think we're going to see it call some tools.

[4:49] You can see it takes a few steps for it to even come up with the actual plan of how to use the right data source here, which is to search Day AI, which will give it all of the meeting recordings, the opportunity metadata, the Slack conversations, and it's all multiplayer. And that data is all private and managed based on everything that people feel comfortable sharing and don't feel comfortable sharing. When I have this connected, it very quickly figures out who I'm talking about. It takes a little bit of time to hunt around, but it's hunting around in the right direction, it's finding what it needs, and my guess is it's going to come back with probably a decent answer.

[5:27] That took a couple minutes, and it's not quite right. […] That's not what's going on here, and this email is not written the way I would write it. And we see em dashes, and it's just a little stilted, it's a little awkward. It doesn't have anything to do with why the deal is actually stuck. It's close, but not quite right.

[5:52] Now, if I use the same customer memory and I do it in the right harness, let's see what happens. This is going to bring us into real brown belt here, that's a word. To really use Fable properly, I'm going to go to the terminal and I'm going to dive into Claude Code, but I'm not going to go in unarmed. I'm going to go in with the right starting point and machinery for Claude Code to really make this happen for me.

[6:16] So I'm going to make a new folder here and say GTM loop. cd GTM loop, and I'm going to run Claude Code. Now, I'm in a completely empty folder that I just created. I have this goal of doing what? Well, I want to find out how to automate every time a deal is stuck where I could actually personally help. Come up with a plan for how I should be involved, write any Slack messages, emails, create email instructions or email templates to make sure it's exactly perfect for that type of situation. Maybe I want to understand all the different ways that deals get stuck. I mean, absolutely massive amounts of work. This is how you're supposed to use Fable. It's very different from how we're used to using Opus and Sonnet and ChatGPT and everything that came before it.

[6:58] I'm in Claude Code, I have an empty folder. First thing is I'm going to do the /model command and make sure I'm using Fable. Perfect. The second thing I'm going to do is I'm going to grab this GTM Brain template that we have on GitHub. This is public. We offer this to people as a starting point. All you have to do is tell Claude Code about it. The link's in the description. I'm going to grab it right here and just say what I want to do: "Let's take a look at this link and set it up with the ultimate goal of figuring out how to identify stuck deals where I, Christopher, the CEO, can help. That's it. I really want to be of service to customers, prospects, and my team, and work my butt off thoughtfully to drive our growth."

[7:48] So it knows exactly how to go get all of the code from this GTM Brain template that we have on GitHub. It's going to pull it all down, and it's going to evaluate what it wants to use. We have a bunch of patterns and starting points and instructions. It's everything that Claude needs to understand how to do a Fable-level project like I'm talking about. I can come into this thing and say anything about what I want to have happen to my go-to-market, and this GTM Brain repo is going to teach Claude Code how to work with me, ask me questions, look at the data inside our workspace, come up with a plan, and actually deploy the right automations with me as the approver, making sure I understand exactly what's going to go on, with ways of evaluating it ongoing. This is just absolutely completely insane.

[8:41] Now, anytime I open up Claude in this folder, I can just run this skill and pick up where I left off: figure out what's going on with my business, what kind of initiatives and plans I'm trying to deploy out to my team and to myself and to my customers, and work with me to drive my business forward.

[9:00] Now, Fable has taken the GTM Brain, connected to Day AI, and learned everything about our company. I can say, "Tell me everything about our company. This is for a demo, so use internal names and roles, but keep customer info redacted." It's really easy for Fable to get all of this primary source information about who's at our company, what our activities are, all the way down to the actual words that we're saying on our calls, and whether somebody's screen sharing, what we're talking about, the pain that prospects come to us trying to solve, everything. This is not just traditional CRM data. This is where Fable really shines. It's not at the individual contributor level. It's like CEO level. It doesn't mean that you can't use it as an individual contributor. Anybody can use it.

[9:48] So what I show you next illustrates just how different working with Fable is from working with the Opus and Sonnet and Haiku models that came before it. When I'm working with Fable, I go up a level and I start to create loops. I create ongoing processes that are aware of how they are doing and able to improve themselves over time. This is a major, major leap forward. It's not about drafting an email. It's about identifying every single moment where I could be helping accelerate a deal and doing the right thing. If that means writing an email, then it is going to figure out all of the information and exactly the right thing for me to say to the right person at that prospect company forever.

[10:29] It'll start running right away. Next week, it'll be running better. The week after that, even better. It may have questions along the way of, "Geez, what do you think about trying this idea?" And I might say, "Yeah, it's worth a shot." It can A/B test itself. I mean, it can do absolutely anything. So I'll just use Wispr Flow and tell it exactly what I want.

[10:48] "Hey Fable, it's Christopher. I would like to design a system here with you that identifies every single place that I could be involved to help accelerate one of our sales deals, take action as quickly as I possibly can with emails or Slacks, any action that best fits the situation. I want you to look into exactly the places that I can help, draft the emails, Slack me, email me, whatever you need to do, send a carrier pigeon if necessary, so that I jump in and help our sales reps close these deals. And I'm not asking you to do this one time right here with me. That's easy for you. I want to give you a real challenge. I want something automated and recurring. And I want to design it in a way so that you can look at your work and improve over time. These agents and skills that we're using here, feel free to suggest new skills and the right automation, Day AI pages to maintain state. Really use your creativity and design something here that you can learn from and optimize yourself over time, both with my feedback and without my feedback."

[11:56] So that pops right into Claude Code. I can hit enter, and it's going to know exactly what to do. And if there's one thing that Fable's good at, it's designing systems, not just for itself, meaning Fable building Fable automation, but especially building systems that use the less expensive, lighter-weight, faster models like Opus and Sonnet. It's super good […]. This is a huge breakthrough in our thinking and also really good news. A lot of people are sour on Fable because it's so expensive, but it really doesn't have to be. You can use a tiny bit of it and use it to get much better results out of Opus, and use Opus to get better results out of Fable.

[12:49] The first thing it's deciding to do is what any good software engineer would do before jumping in and adding a feature to some product it just started working on. It's going to see all of the automation, the agents I have working for me already, their roles, responsibilities, things they're doing automatically on cadences and in response to particular events. Taking that into account, it's going to make a decision of maybe to improve one thing over here, borrow from something over here, add a skill to an existing agent, maybe reword the job description of my chief of staff agent. I'm really curious to see what it comes up with, but it has a big canvas, mostly blank, but there's some stuff on it, and it's taking all of that into account.

[13:29] Super smart. This gives me a lot of confidence that I can do ambitious things and run big scary ideas by it, and that it's not going to just run off and ruin my workspace. It's going to be very thoughtful about this, and it's going to present these options to me and work with me to build this system piece by piece. The GTM Brain that we publish at Day AI: Claude has decided to borrow some of these skills and use them. It's not something you have to use. You can build one of these GTM brains yourself. […] It doesn't lock you into anything. It allows and encourages you to bring your own spin and perspective to the process.

[14:17] I do a lot of this kind of stuff, and this is the part where, to be honest, I'll probably go get a cup of coffee or something, or an energy drink, or maybe chip away at another task or have a conversation around the office and let it run. I have Fable stuff. When it is building a system, it might run for 10 minutes. It might run for half an hour, depending on what it's trying to accomplish and how many subagents it's deployed, and that's fine, because it is building me future leverage, okay? It's a little bit like creating the house that I'm going to live in and move through every day without thinking about it.

[14:54] One eternity later, it returns. Very cool. It has created an initiative called CEO Unstick Deals. It's doing this as a record of the project that we want to do. It's not just going to go update my workspace and start doing all this stuff. It's going to talk to me about it a little bit. It's going to update its plan. This is the go-to-market version of what we do in software engineering. Everybody who is using AI in software engineering has some three-step process that starts with planning, agreeing on the plan in detail, implementation, and then verification, testing, acceptance, ship it. And the human can be involved in different ways across these three steps.

[15:31] And it's really different from 2 years ago, where the AI was kind of chipping in on the implementation. Now the implementation happens and the software engineer usually just never even sees it. We're dealing with the intent and the details of the plan, and we are dealing with the result: the taste, security concerns, anything around accepting the final result. And that middle part has pretty much gone away. We're going to do a similar thing here. This looks really good. Let's go ahead and deploy it to production. And it chomps away a little bit and pushes this up to the cloud, totally securely, on the right guardrails.

[16:04] And here we are. "All right, Christopher, the CEO assist system is deployed and live. No carrier pigeons needed. From Monday morning, the places you can accelerate a deal come to you with the move already drafted." So here's what's running now. Every weekday at 7:15, my chief of staff-y agent, Abby, is going to sweep every open deal across three pipelines that we have, and it's going to look for six different moments. A deal that's gone quiet, close dates that have slipped, named blockers on the opportunity. So the AI has already picked up that there is a security process or budget scrutiny, etc. It's going to look for all of those. If there's a decision moment with a CEO I haven't talked to yet, anything waiting on me personally. I hadn't thought of that. That would never happen. And it says, "Okay, we're going to look for those first." Okay, maybe I missed a DocuSign for a mutual NDA. God forbid. And I like this idea. It says, "Okay, on closed won or closed lost, where I should reach out with a note from the founder and kind of connect with the prospect, whether it goes our way or not."

[17:08] It's going to take those and then judge for each of those which ones matter. Is it something that specifically benefits from me being involved? I mean, this is all the stuff that my team is doing. Probably spending minutes, hours a day, each of them, deciding whether to bug me with something or packaging it up for me. They at least know that they can point me in the right direction and I will have all the context from Day AI, but there's still some cognitive load on, you know, is this worth bothering Christopher? I'm going to remove all of that. I'm just going to hop in on all of these things in the way that the reps want me to without them having to even ask. How cool is that?

[17:47] Okay, and it's going to rank. This is a great trick with having Claude, especially Fable, do any kind of creative work (and I consider this very much creative work), is to generate a bunch of options, a bunch of places that it could go do work for you, and then have a way for it to rank those so that it's not just spinning off into eternity, that it prioritizes its own work. That's incredible. That's another piece of working with the AI that is now taken off my plate. So much over the last 2-3 years of working with these serious AI systems has frankly been us sitting there deciding what to ask Claude to do and not. We have a million things. We have a huge backlog. Let's give it a way, a rubric, to decide what's worth working on and what isn't.

[18:29] And it's going to give me the evidence, it's going to give me the background, it's going to draft the email for me based on all of that research. And then there are certain circumstances here where it's going to just be left to the rep. I might get notified, I might not. Wonderful. All right, it's going to let me know whenever there is progress in a deal that I might want to pile on to. Whenever there's momentum that I can add a little bit of flavor to, get involved a little bit, it's going to let me know about those. And then on Fridays, see, this is the key part. […]

[19:03] You see how I asked it to go to work? From the very beginning of how I spoke to Fable, I said, "We're going to [build] a system, okay? We're going to architect and implement a loop." And I was very careful to really underscore the importance that it design a system that it could not just do the work, but evaluate the quality of the work and improve all of the logic and all of the instructions and context of that loop as it runs. So the more these tokens process, the more these events fire, the more I weigh in on it, the better it gets without me doing anything.

[19:40] This is very difficult to build with any model that came before Opus, and it will completely change, or it should completely change, the way that you approach working with AI at work. This is nuts, and it's telling me here how it's going to do that. It's made a couple of assets that live in the cloud. There are Day AI pages. These are like really nice, clean wiki pages. I can make them private to me and my agents. I can expose them to the rest of the company. I can, frankly, never open them at all. I can actually make them available to other people's agents.

[20:09] So now I can create a system for myself and have part of that system hook into somebody else's system. I can create loops around our entire go-to-market, where every meeting, every Slack is automatically looking for opportunities to up-level our game, pulling us in from leadership where we can weigh in or make a decision, and simultaneously pushing the right coaching and the right tips to all of the people on the front lines interacting with customers in that first call, or transitioning from sales to engineering after the purchase. Everything can be automated. Everything can be a loop. Everything can be improved dynamically every time it runs.

[20:49] And then, the icing on the cake here is it gave me a command in Claude Code that I can use to evaluate how it is evaluating itself, which is just wild. This is that same idea of going up a level. Once you've gone up a level from one unit of work to all of the units of work of that type, all of the situations that cause me to have to do something like write an email to a CEO, I can go up another level and say, "Is this the right process? What are the other things like this that I could be doing?" Let's say I build some of those out. How are all of those working? How could those work better together? What could David be doing? What could Laura be doing? I can push all of this stuff out. We can sit around and talk about all this stuff, take our recording of ideas and come up with plans and push it all out there, and just improve this system every single day just by having ideas, getting the data, seeing the feedback, hearing from people at the water cooler. This is unbelievable.

[21:48] All right, so I hope that gives you a really good idea of just how different using Fable in the way that Anthropic is saying that we should. It's a big up-leveling of our thinking, but it's not hard to do at all. And honestly, now, even in the last week or two, I'm starting to think about this stuff very differently. Instead of just doing a piece of work, I'm finding myself coming in and saying, "Build me a system that…" I'm having a ton of fun with this, and I hope you do, too.