[0:00]Most people's AI agents make decisions like an intern on their first day. They grab the first idea that comes to mind. They patch symptoms instead of actually finding the root cause and fixing it.
[0:11]They trade quality for speed without even thinking about it. And then you wonder why you can't trust them with anything important.
[0:17]My name is Andrew. I run 11 AI agents across three businesses, two of which are seven figure companies. I've logged over 700 hours building inside of Openclaw at this point and I took my AI agency to multiple five with a single human employee.
[0:34]This is the open cloud blueprint series. This is video 4 of 8 in the Openclaw blueprint series and today we are covering the principles I programmed into my agents that make them think like a CEO instead of a simple-minded robot.
[0:49]Now a quick question. Have you ever asked your agent for a solution and it gave you the first idea that came to mind completely skipping over better options? Yeah, me too.
[0:59]And have you ever watched your agent fix the same bug three different ways across three different sessions, never actually solving the underlying problem.
[1:10]Of course you have, because I've experienced that as well. And have you ever wondered why your agent treats every decision like it's the same level of importance instead of weighing what actually matters most.
[1:21]Now I'm not going to be lying to you. This one, this one drove me the craziest. Now it's not that your agent is dumb. It's that you never taught it how to think. And more importantly, most people don't even can think about teaching their AI agents how to think.
[1:34]Well, in this video, I'm going to show you how I taught my agents to think and you can take from this video what you want, utilize it and make it your own however you want to.
[1:45]But the principle remains the same, which is you need to teach your AI agent how to think.
[1:50]Now here's the moment I figured this out. All right, I'm not going to lie. I had Gangas Khan, my CEO agent uh building a feature for one of my businesses.
[1:58]He shipped it and it seemed to work and I just moved on and was like, all right, cool, this works, right? And then three days later, the feature broke in a way that affected my client's operations.
[2:06]He fixed the immediate symptom and then two days after that, it just broke again and a different symptom and the same underlying case and cause though.
[2:13]And everybody fixed it and then, you know, it broke a third time and I'm like, this is ridiculous. You know, that's when I actually realized that he wasn't actually solving anything. He was just patching every single time.
[2:24]And he was doing it because nobody ever told him that there was a difference between fixing a symptom and then fixing a root cause to a problem.
[2:32]Now, let me tell you why this matters at scale, okay? If you're running one agent on a simple tasks, mediocre decision making is annoying, but, you know, it's intolerable.
[2:42]But if you're running 11 agents like myself, handling real business operations, mediocre decision making compounds into absolute chaos, all right?
[2:48]Bad decisions end up snowballing and then root problems never even get solved. The system you thought was saving you time, it becomes a system you're constantly cleaning up after and having to babysit.
[2:59]And I wasn't going to accept that. So I sat down and I wrote out the six principles that I actually used to make decisions in my businesses and I baked them into a single file that every one of my agents reads, the thinking.md file.
[3:13]And it changed everything for me. Now, if you guys are getting value out of this video and you want to connect with me directly and join a private live that I do inside of Telegram,
[3:20]then go ahead and join that Telegram channel. The link is down in the description below. It's where people are going beginners and business owners and operators to go ahead and talk to me in a Telegram live setting where you can just sit there and chat with me about anything that you have questions on.
[3:37]It's where you can also influence the type of content that I make on this channel and what I cover in my live streams every Tuesday and Thursday.
[3:45]And the videos come out every Monday, Wednesday and Friday at 8:00 a.m. So, if you want to be able to have a level of influence over what I cover and have that considered because it's something that you need help with, probably along with others,
[3:55]then go ahead and click the link down in the description below. It's completely free and I have absolutely nothing to sell you in there.
[4:02]Now, most YouTubers teach you prompts, okay? They teach you the tricks to get better outputs from your agents.
[4:08]That's extremely surface level if I'm being honest. Real operators don't optimize prompts, okay?
[4:14]They program decision making frameworks into their agents and the thinking.md is the file where that all happens.
[4:22]Now, I'm going to walk you through all six principles. Each one solves a specific failure mode.
[4:29]I was watching my agents hit personally and each one is a non-negotiable in my system. Principle number one is map out all the options first, okay?
[4:36]When you ask most agents for a solution, they give you the first idea that comes to mind. It's not decision making, that's just a knee jerk reaction or a reflex.
[4:44]But the principle I program in is this. Okay, before recommending anything, the agent maps every possible option.
[4:52]Plays out each one and it evaluates it on one of three axes, effectiveness, efficiency, and sustainability.
[5:01]And then it presents the best path with the tradeoffs very clear. Now, I also have the agent tell me all of the options that it thought of and why it chose this option over the other ones.
[5:14]Now, the rule that locks this in is when effectiveness and efficiency conflict. Effectiveness always wins. Always.
[5:22]I'd rather pay more for an outcome that actually works than save money on a solution that doesn't and I have to keep patching every single time. It's just not worth it to me.
[5:32]This principle alone eliminates probably half the bad decisions against make. Principle number two is fix root causes not symptoms.
[5:40]Standard agent behavior is something breaks, the agent fixes the immediate problem and then moves on. And the same issue breaks again next week.
[5:49]Different fix, same underlying cause. endless fire fighting and a bottle neck and just a time suck.
[5:56]The principle I program in is that the break fix protocol. Okay, when something break the agent peels back layers until it finds the actual root cause.
[6:05]Then it asks what's the cascading effect? What else did this break or what else could this break?
[6:10]Then it fixes the root issue and then sweeps for the same gap elsewhere in the systems.
[6:16]And also note that I have the agent tell me what the root problem was and why this happened, what caused it and this is important to notice as well for future reference.
[6:24]Now a real example from one of my agents is when Mercury scheduler was breaking daily.
[6:29]Surface problem looked like it's an API timeout, right? Layer one was the API was timing out.
[6:35]Layer two was too many concurrent slack channels being checked. And layer three is the agent was pulling 65 individual slack channels every single cycle.
[6:45]And the actual route cause was the architecture itself. Now, Mercury should have been monitoring the channels for clients who are either marked as next round eligible or monitoring since those are the channels she is monitoring for a response from her previous message.
[6:58]Once we fixed the route, the symptom never came back. And we found three other agents doing similar inefficient polling that we've then cleaned up at the exact same time with this.
[7:09]Now, principle number three is the priority filter. Essentially, what eats my time up. Now, most agents treat every task as equally important and that is completely wrong, okay?
[7:19]Some failures are catastrophic, but some are annoying and some are barely worth even touching at that point in time.
[7:25]So, the hierarchy I program in is system and infrastructure health is priority number one. API keys, gateway connectivity, and agent communication are the foundation everything runs on.
[7:35]They get fixed immediately. Client facing operations are priority number two. And anything that touches a real client's experience is next.
[7:44]Internal operations are priority number three and they're important, but they're not super urgent. The meta rule is simple. Weigh out everything by what consistently eats my time.
[7:54]If a system breaks daily and I have to manually fix it, that's the highest priority regardless of the category because every minute I'm fixing this infrastructure and all these issues is a minute that I am not building the business or focused on revenue generating tasks.
[8:07]Now, really quick, if you guys are gaining some value from this video and you guys want to connect with me directly,
[8:15]then you can go ahead and click the link down in the description below in order to follow me on X. A lot of people are reaching out to me there directly and sending me DM requests for specific scenarios and just things that they're facing as issues that they just may not want to put down in the comment section below.
[8:30]So if that's you and you want to go ahead and ask me any questions, you can shoot me a DM there or if you just want to follow me on X to just, you know, keep up with everything I have going on,
[8:40]then go ahead and click the link down in the description below to give me a follow. Now, principle number four is resource philosophy, okay?
[8:47]Main thing, protect the base. Every business has a stable revenue base that everything else depends on.
[8:54]The principle I program in is simple. Every minute the agent freeze up by automating something is a minute I get back for revenue generating task.
[9:03]Whether that's putting on new clients, getting new offers out, strategic decisions, right?
[9:08]The growth activities that actually scale the business and move the things forward. So agents aren't automating just for the sake of automating, right?
[9:16]They're automating to protect my finite resources, my time, my energy, my head space, like when Mercury scheduler breaks daily, that's not just a bug, okay?
[9:23]That's stealing time from growth activities and which makes it much more expensive in a way of the bug itself doesn't capture.
[9:30]Now this principle changes how the agent prioritizes. It stops asking what's broken right now and it then starts asking what's costing Andrew the most time right now that he has to keep spending his time on.
[9:43]What's costing him the most? I'm going to go solve that. That is why principle number four is super, super important. Now principle number five is the done standard, okay?
[9:52]I cover this briefly in video number one, but it's worth repeating because this is one of the principles that changed the entire system for me for the better.
[10:02]Okay, now most agents define done as, I built it. It seems to work. Let's go ahead and ship it. That's not done to me, okay?
[10:09]That's halfway done if I'm going to be completely and utterly honest with you. My definition of done that is programmed into every agent's thinking. D file is done means Andrew does not have to think about it anymore, okay?
[10:22]If I'm still watching this system, still worrying about it and still manually checking, it's not done, okay?
[10:29]If the agents out put still requires me to review every single time it gets something done, that agent is not actually autonomous, okay?
[10:40]It's just a task runner with a couple extra steps and then I'm a task runner that I quite frankly have to baby sit and I don't want to. The test I program in though to validate whether or not this principle is valid is very simple.
[10:49]Can this run for a full week without me even having to look at it? If yes, then it's done. If no, then it definitely needs more work.
[10:57]Then I have the agent go back in and understand why this is still not working flawlessly for a week straight and tell me all of the reasons and potential solutions to this issue.
[11:07]Now nothing's ever going to be 100% perfect, but I'd like it to operate at 95%, you know, uh clean like clean runs and no issues, right?
[11:16]You know, this one principle though is worth probably 10 hours of my time per week because it forces every agent to ask itself a higher question before declaring something as completely quote-unquote done.
[11:26]Now for principle number six, this one's a good one as well. And this is input over output. Now, the standard debugging behavior is very simple, all right?
[11:34]Check if the output looks right, if it's wrong, tweak the settings, try again, and it's just reactive problem solving and the agent is always one step behind.
[11:44]I hate that. I hate reactive problem solving. The principle I program in is the opposite, okay?
[11:51]Don't watch the output so much. Watch the input through every step because perfect input generates and guarantees perfect output.
[12:00]And if you obsess over input quality, output becomes consistently high quality without you ever having to fix it on the back end.
[12:06]Now, a real example of this is when Mercury's booking system wasn't working. The old way of thinking is just to fix the bad bookings, but the new way of thinking traces the inputs, okay?
[12:17]So step one, did the client message arrive correctly? Yes, okay, then step two, was the availability parts correctly? No, it was not. Okay, there's the problem.
[12:27]The client said, let's say after 9:30 and the system was treating it as at 9:30 only.
[12:33]And now we found the bug in step two of the input chain before it ever became a bad output again. Now, this principle teaches the agent to investigate forward, not patch backwards.
[12:46]Okay, this is very, very important. It's the difference between like a junior employee and a senior one. And don't get me wrong. I do have an agent that still monitors the output of all of my agents that are actually operating in my client's businesses.
[12:58]So let's not act like I'm just completely ignoring the output. I also have an agent doing that, but I also want to ingrain this framework and this principle for the agent to operate upon all the time, no matter what.
[13:11]So that's the system, okay? Six principles, map all options out, fix root causes, filter by priority, protect the base, define done correctly and watch input over output.
[13:24]Now, these aren't suggestions, okay? I'm just going to be completely transparent with you. Like, they're not preferences either. Like, they live in the thing.md and they get loaded every single session and they get reloaded after every context compaction as part of the nine file protocol from video number three.
[13:40]Which, if you did miss video number three, the link is down in the description below for that. And every decision my agent makes runs through these six principles before any output ever reaches me, okay?
[13:52]That is the difference between an agent that thinks like an actual CEO and an agent that thinks like an intern.
[14:00]The intern just simply reacts to what's happening, but the CEO has a framework and is completely and utterly proactive.
[14:07]Now, let me show you exactly how to set this up because principles you don't implement are honestly just really good ideas at that point.
[14:14]Now, this entire framework lives on a file called thinking.md as I've mentioned before. Each one of your agents has its own thinking.md file in their workspace.
[14:24]The same six principles, same exact structure, same standards across every agent in your operation. That consistency is the whole point.
[14:32]Now, the file structure is very, very simple to follow, okay? Six H2 headers. Okay, one for each principle.
[14:39]Under each header, the rule itself in plain English. And then under that, the reason behind it.
[14:47]And then under that, the test the agent should run when applying this principle. Now, the key is that thinking.md isn't a reference file, okay?
[14:56]It's an active framework. Your agent should be running every major decision through these six principles before doing anything, okay?
[15:05]Which means in your agent's.md file, you also need a rule that requires the agent to consult the thinking.md before any significant decision.
[15:12]We covered agents.md in video number one and post compaction reload from video number three, so it already pulls the thinking.md back into context after every single compaction, okay?
[15:23]So the system is indeed layered, okay? The thinking.md holds the principles. The agents.md the use of them.
[15:33]And the reload protocol makes sure that they survive every single session. I'm sure you're probably starting to see how I'm culminating all of this and it's all coming together.
[15:41]Just wait, we still have four more videos to go and I have much more value to give in this blueprint series.
[15:47]Now for the biggest mistake people make, okay? They write thinking.md files once and then never reference it again.
[15:54]The agent reads it on day one, then it drifts away from it and that's why agent.md enforcement rules actually matter.
[16:02]Without it, the principles are just decoration at that point. Now, I'll be honest with you. Once I locked in this thinking.md across all 11 agents,
[16:11]the change was immediate, it was definitely obvious, okay? My agents stopped patching symptoms, they started actually solving problems.
[16:20]And they stopped treating every task as equally important. They started filtering by what actually matters.
[16:26]And the quality of decisions across my entire operations jumped tremendously. Didn't go from nothing to perfect, but it improved tremendously.
[16:34]Gangas Khan, my CEO agent, absolute G, he is, started catching root causes before they cascaded.
[16:40]Mercury started watching input chains instead of fixing bad outputs after the fact. And Machiavelli, my strategist, and my right hand man, he started mapping out options properly instead of just recommending the first thing that he thought of.
[16:53]Now, for you, that means agents that actually solve problems instead of patching them. Agents that prioritize correctly without you having to micromanage them and babysit them all the time.
[17:03]Agents that hand you finished work, not work that needs another round of review and the bigger your operation gets, the more this matters.
[17:13]Now, this is the kind of edge that's only available right now, while most operators are still figuring out the prompts, the ones who program decision frameworks like this into their agents are going to scale, sustainably and have less issues and just be overall, making more progress much faster than you are.
[17:32]So, now you understand the six principles. You know the file they live in, you know how to enforce them across your agents, but there's a layer above the principles, hard rules that can never be broken under any circumstance.
[17:48]And in video number five, I'm going to be breaking down the 10 red lines, okay? These are non-negotiable rules that survive every context compaction, every long session and every edge case.
[17:59]If thinking.md is how my agents make decisions, then the red lines are the guard rails that keep them from ever crossing certain boundaries. And this is the type of file that prevents agents from making catastrophic mistakes that could do real damage to you and your business or your clients's businesses if they're using you for front facing client interactions.
[18:19]And if you missed any of the previous videos, watch these first, okay? This series compounds each video makes the next one much more powerful and the link to the playlist is down in the description below.
[18:30]Now, three things before you go as always, number one, subscribe and hit the bell notification so that way you don't miss the rest of the Openclaw Blueprint series.
[18:39]All eight videos, you're getting the absolute full system, okay? Number two is join the Telegram, okay? The link is down in the description below.
[18:45]Business owners and operators and beginners are all hanging out in there, coming and joining the live and influencing exactly the type of content like this that I'm making and putting on this channel.
[18:54]So if you want that, go ahead and click the link down in the description below. And number three is go ahead and give me a follow on X, okay?
[19:00]This is the best way to communicate with me directly and get direct help on your specific situation and get some feedback from me. So, the link for that is also down in the description below.
[19:10]Lastly, this is the Openclaw Blueprint series. I will see you guys in video number five. Peace.
