Transcription
I have four AI executives running a total of 22 agents that control different parts of my business. They work 24/7. They don't call in sick and they coordinate automatically. But here's the thing no one's talking about. Without the proper cron jobs in place, your system falls apart.
So today, I'm going to give you seven prompts that you can use to keep your multi-agent team alive, healthy, and working 24/7. So without squirrels, let's build. All right. As you can see here on screen, I had my executive assistant prep this visual aid. Let's go to number one first, right? Our chief heartbeat monitor.
Now, you want to create one of these for each of your agents. Now, the intervals will be different. The rules will be different. This is why my recommendation out of the gate for all of the prompts I'm about to show you. Don't just copy paste them outright. What you want to do is screenshot this page here, feed it to your agent, and say, "Hey, I'd like to set up heartbeats based off of this example for each of our individual agents," and then you communicate back and forth based on your environment, based on what you have your agents doing on a daily or weekly basis. You configure things that's best for your environment. This is meant to be an example, right? So, this is Elon's config here, right? He his ID is Elon. This is his workspace. This is his interval for the heartbeat. He only works between certain hours that I specify. And then he uses a very cheap and efficient model because you don't need high reasoning for the heartbeat checker. Right? That's what I've learned.
All right. Then moving along. Right. This is what the heartbeat MD looks like. It has every heartbeat. It reads our shared team memory markdown file. It's the unified team brain that I've configured. It then reads its own memory and backlog. And then it checks notes for pending items. And then it'll scan the PM board for teams tasks. This is my where it's hosted, right? Your environment might be a little bit different, but the PM board for me is the single source of truth. So the reason I configured these heartbeats was sometimes I manually go in and move things from the review column to the done column because the review column in my PM board is for things for me to approve, right? I'm that blocker. I'm that stopper that I don't just want them doing things autonomously yet. I hold their hand for certain automations and certain processes before I'm confident and then I create an automated workflow for that particular process. But until I'm happy with it, things go in the review column. Well, an issue I was running into was sometimes I would drag over and mark things as done, but they wouldn't they'd be working off of stale data. So having you know these rules in place on the heartbeat means that every interval that I specify they then update their memory. They update the backlog and they're aware of what's done versus what's in progress versus what's still in review. Right? So this one's very critical. It's very valuable.
All right. So prompt number two. You're going to notice a theme with this one here. This is the only prompt I'm giving you today that technically doesn't have a skill file. Ignore this title here. Okay. This is technically a script. Now the purpose of the watchdog pattern is for your critical crons that you schedule that are responsible for in this example our daily executive syncs. This is where our agents come together and they discuss right we have certain we're going to show that executive sync a little bit later in this demo but sometimes they might fail. So we have this watchdog in place to first verify that that cron completed. If it didn't complete, it runs this script right now. Copy paste this. I'm gonna I'm gonna sound like a broken record here. You can't just copy paste this and tell your agent to incorporate this line for line into your system. My system's unique. The models I use may not be the same models that you use. The names of my crons are not the same names as your crons. The paths of how I have my system set up are completely different than your paths. So, it's my recommendation that you screenshot this, feed it to your agent and say, "Hey, I'd like to follow this example and create a watchdog pattern for each of our most critical crons that we have in place." You have to decide which ones you want to apply a watchdog cron to. One other thing important thing to note here, right, our daily executive syncs that requires high reasoning. So, I do use Claudius 46. However, right, it's just referencing that critical job. I use Haiku for this particular watchdog because it doesn't need a high reasoning model. It's running a script that never changes, right? So, it's just a little trick. I have a feeling you're going to find some value in that one.
Now, moving along to our cross-functional executive sync. This is the fun one. This is the one that runs every single day. It spins up each individual agent and they have a conversation autonomously without me. I get the transcript. I get the audio playback at the end of this. But what I want you to do if you want to create your own for your own system, screenshot this page here. Okay, this is what my cron looks like for the daily executive sync. I have it run Opus 46 of course, right? It's an isolated session and I do have a message, right? Just letting Muddy know, hey, this is the daily executive sync. And then I have it read and follow the executive sync skill MD. I do have a few rules in here, but these are mainly environment-based rules.
Now, moving along to the executive sync skill, right? I have a step zero. So, this is a pre-meeting context load. First, it pulls from the PM board, then the suite chat if I write any notes in there, and then previous transcripts, right? So, this is for anti-repetition. Then, I I pull live YouTube stats since I have them discuss our channel every single day. Uh, then we have a shared context markdown file and a day of the week tailored agenda. Step number one, it only surfaces items with status changes. It skips unchanged items entirely. Step number two, it'll spawn each individual chief. Right? These are the real agents that all run in parallel. I can see it in my active sessions. I can see each individual agent actually responding to one another. It's pretty nifty. We also just showed you how to build this system in yesterday's boot camp. It was almost a five-hour raw live stream where you see the step-by-step of me setting up a multi-agent team with stand-up features just like this. Now, and then it, you know, it has a step three and then a step four, right? I use an open-source uh solution. You can use 11 Labs, whatever you'd like. But what it does is it saves, it transcribes the audio, it updates the PM board, and then it pings me on Telegram with the executive sync after every morning meeting. This is also very similar to my weekly executive syncs as well.
Moving along to nightly build and deploy. This one's fun. This is a self-improvement. As I mentioned before, screenshot this page if you want to feed it to your agent to create your own version of it. Now, this one, this example is specifically for the nightly clear mud OS, right? I do have it running Opus 46. It does need Hydro. I've, you know, I've tested back and forth with different models. I've just found that I'm happier with the output. I have less critique, less feedback. I'm generally nine times out of 10 going to keep the feature and happy with the self-improvement that it made overnight. Right? And you're going to see I just titled the message, right? uh I just title the cron and then I have it read and follow this skill MD and then I do have some environment-specific details in the the cron as well. Now this is what my skill MD looks like. Step one, we select first we check the backlog, then we check bug fixes second and then we polish existing features third. If one through three have nothing to add, right? Then it will create new features. This is a workflow I had to to develop over time. One thing I found with my first iterations of my nightly build was it was building me new features, but then it started to get redundant. It actually wasn't checking the entire codebase. It wasn't checking if there's something similar that could be improved on. So, this is something that I had to kind of go through day by day, week by week to fine-tune. And hey, maybe you don't need that step one, but this is just meant to be an example, okay? And I'm going to show you what that section looks like here in a moment.
All right, moving along to workspace memory sync. Now, this one's very critical to make sure there's no stale data, to make sure all of your agents are making sure that they're all on the same page because that doesn't technically happen real-time. Now, this is what my cron looks like labeled chief workspace sync. I have it read this skill MD off to the right and then I only have it running Sonnet 46 because it doesn't need high reasoning. It doesn't need Opus's capabilities for this particular automation. Right? This is just more of a synchronization than it is a task.
All right. So step one it reads all of these sources. Step two is it checks the PM board. This is critical. It compares each of the uh executives memory MD against the PM board status. Any item marked active in memory but done on the PM board must be updated to resolve. Now, this is something critical I had to incorporate because when I manually go in and do things at the end of the day on the PM board, there was a disconnect. Agents wouldn't then get synchronized with that information. So, the morning after I would receive meeting notes saying, you know, this hasn't been done yet. This has been stale for X amount of days when I already went out. I already responded to the lead. I already made the fix. Whatever the task is, there was a clear disconnect. So, this was a step 1B that I had to incorporate. And ever since I incorporated this step, it has been solid. The PM board is that single source of truth. You get the picture here. So, I can't stress this enough. Screenshot this, feed it to your agent, and say, "Hey, I'd like to build something similar. What can we do based off of our infrastructure and our agents and our, you know, our workflows?" Right? Because the way your system works is not the way my system works. So you can't just copy this outright. What you have to do is screenshot this, feed it to your agent, say, "Hey, how can we incorporate this into our system?"
Moving along to Saturday Vision Sync. This is a recent one that I incorporated because I'm a big vision guy. I like to visualize the future. I like to plan ahead. I like to think in depth. And ever since I started these daily and weekly executive syncs, well, they were handling my day-to-day and week to week, but they're not thinking big picture, right? So, I thought, wow, let me create an executive sync that happens normally on their day off, which now they work every day of the week. Uh, I used to give them Saturdays off for my infra day to upgrade the system, take it offline if I need to, etc., etc. But now they also do these Saturday vision syncs where they think big picture, right? They don't talk day-to-day. They don't talk tasks. They don't talk what was completed the past week. They reference the user MD and they think big picture, right? So this is what the cron looks like. And you're going to notice most all my crons always have a skill tied to them. I believe all but one of the crons we show off in this video have a skill tied to them. And you're more than welcome to screenshot this, feed it to your agent and say, "Hey, I'd like to incorporate something similar into our system." Right? One thing I can't stress enough, right, is if you copy paste this and tell your agent to follow it line for line, well, your environment is different than my environment. So that's not going to work out right. You want to use this as inspiration.
Now, all right. And lastly, memory maintenance and pruning. So this one's critical. This one has prevented me from having to explore QMD. All these memory related solutions that people are deploying and then all of these memory issues that people are complaining about. Listen, I've had some issues and this is why I created this cron plus skill MD to help maintain memory. And I'm going to be honest like there's the occasional drift, but for the most part 9.9 out of 10 times I have zero memory issues. It is freaking amazing and it's all thanks to this cron and skill pairing. So, screenshot this, feed it to your agent, and apply it to your system. Okay? And we'll we'll go over the skill here briefly.
All right? So, let's just review the skill really quickly, right? Step number one is it reviews the past week's daily files. Step number two, it extracts significant items, decisions made, lessons learned, new capabilities, milestones, etc. It skips the routine, the backups, the heartbeats, the normal crons. Then it updates the memory MD, right? It adds distilled learnings, removes stale entries, and consolidates duplicates. This is capped at 500 lines. I'm very aggressive about about pruning. So I have this in my skill. Now step number four is it reconciles the backlog MD. Done but not checked off, mark complete. No longer relevant, remove with note. Stale over two weeks, flag it. Steps number five to six is gap detection plus chief cross-check. This verifies daily file exists for each of seven days. It scans chief workspaces for info missing from our central memory MD. And then step seven right this is the cron prompt drift audit. It runs a Python script to check all cron jobs for bloated inline prompts over 20 lines no skill reference. Right. So, this is something that I have that it's like a checker each week that if something slips through the cracks and I have too much information in a cron, it flags that cron and lets me know and then I can address it, right? And then step number eight, it logs all maintenance results, files reviewed, items added, removed, flagged, and yeah, so feed this to your agent. I have a feeling it's going to improve your system. And yeah, I hope you found some value in today's video.
Now, I'm not an AI expert. I'm building in public and sharing what actually works. If you'd like to see me make a particular video around a specific topic, drop a comment below. Let me know who you are, what you do, who your target audience is, and what your question is, and I will add a video to my queue custom-tailored just for you. Thank you so much for tuning in today's video. My name is Marcelo. This is Clear Mud, and clarity matters.