AI agent 18 years ago
脈報 · 2026-06-02 · 10 min
Show full episode description
多代理 AI 老是互相踩線、崩潰全毀？這位 YouTuber 用一塊共享看板當協調層，讓最多 18 個 AI agent 平行接力做研究、判斷、產出，整條流程只留一個人類核可閘門。 ⭐ 文章深度讀：我把「協調為什麼才是多代理真正的瓶頸，以及這套看板設計哪裡值得抄、哪裡只是噱頭」整理進文章，用脈報的角度拆得比影片更白。 → https://heymaibao.com/multi-agent-kanban-workflow/ ⚡ 章節重點 開場：多代理為什麼老是壞掉
📑 Chapters — tap a time to jump there
00:00
核心結論：別讓 agent 互相對話
00:52
看板就是單一真相 (一個 SQLite 檔)
02:17
Dispatcher：一張卡怎麼變成 agent
02:58
事件驅動依賴圖與五大優勢
03:39
實測：獵取 AI 痛點的流水線
06:48
唯一的人類閘門與自我修復
07:45
重點收穫與結語
🌐 This transcript was automatically translated to English from the original.
Welcome to this visual analysis. If you are like me, you are a content creator or knowledge worker who wants to use AI to crazy improve productivity, then you must have heard about how amazing multi-agent AI is. Imagine a whole group of AIs helping you find information, write articles, and even directly extend the finished product at the same time. It sounds like magic, right? But if you really roll up your sleeves and run, you will know how skinny the reality is. What is the most common end result? There is still a budget to burn, and then as long as one of the AIs suddenly crashes, wow, the entire project progress will be wiped out. So today, I will share with you a super pragmatic and even counter-intuitive major discovery. How to make this group of AIs cooperate obediently, and never take the opportunity. Let’s first reveal the most important conclusion, which is today’s trump card. When building a multi-agent interaction flow, the most common fatal mistake everyone makes is trying to let the AI agents talk to each other. You may be thinking Human teams rely on communication to cooperate, but when it comes to AI, because they don’t have a shared progress memory, they will scurry around like headless flies when they start chatting. Letting them talk to each other or rely on that kind of complex message transmission will only make the system fall into an endless dead loop and burn your computing resources in vain. The author of this material proposed a very direct solution. Just cut off their conversation and let them all work on the same board. Okay, let’s compare these two approaches directly. You can see on the left is the model that everyone used to do in the past. Agents chat with each other or send messages to communicate status. On the surface, it looks very cool and advanced, but in the end, it almost always hits the same iceberg, which is a coordination disaster. As soon as the system becomes complex, if any link in the middle fails, the entire progress will be lost. But look at the right. This is the savior we are talking about today. The single source of truth, which is the shared dashboard. This new method fundamentally solves the chaos. It turns the originally extremely fragile multi-agent collaboration into a super stable system. So what kind of magic technology is this bulletin board? To put it bluntly, aside from those gorgeous AI coatings, its bottom layer is actually just an extremely simple SQL Lake database file. This bulletin board is the only truth on the table. Under this architecture, there is absolutely zero dialogue between AI agents, and there is no complicated information to help. They only do two things every day. First, read this board. Second, update the status of the card that was photographed according to their own configuration files. He also plays the role of a message hub and an accumulation log. Minimize unnecessary noise directly. In order to make everyone understand better, you can imagine that this is a whiteboard covered with sticky notes. Let's dismantle how this engine works. The first step is that there is a card on the board, which changes to the old sequence. Then the second step is that there is a loop called the scheduler, which is the Dispatcher. Claim this card. This action is super critical. It ensures that there will never be two AIs running to grab the same job. Then the third step is in a clean and independent work area. The assigned AI agent starts to execute the stirring items on this card. Finally, the task is completed. The AI marks the card as completed, and the circle continues to rotate. Because all states are firmly remembered on the board, even if the system suddenly restarts, it can continue perfectly, and the stick will never be dropped. Next, we enter the first part. Event-driven dependencies. Tasks will be automatically queued to wait. The smartest part of this design is that the sub-task will stay in the agent state until his side task is completed, and he will automatically upgrade his status to ready. The whole process does not require you to write a bunch of messy communication codes for connection. It is very elegant. Let’s quickly sort out the five core advantages of this architecture. First, it is really very durable. All states are written into the database. You don’t have to worry about restarting or crashing. Second, parallel processing. You can release a lot of agents to run at the same time. They will never fight with each other. Third, event inactivity, the workflow will go down very smoothly and naturally. Fourth, it actually has self-healing ability. If a certain task dies halfway through, the system will automatically recycle it and regenerate it. Finally, it is completely collectible. No matter which time it is claimed, executed or completed, there are complete records for checking. These five points added together can be said to completely break an assembly line that was originally broken at a touch. This is not a theory on paper. The author of the data personally went to build this assembly line and used an extremely pragmatic automated factory to directly prove to us that this structure can really be built. Let's take a look at the super crazy staff of this factory. The outermost layer is the red queen. They are like scouts responsible for exploring the outside. One of the red queens is equipped with a Grox model to scan the X platform, while the other is responsible for patrolling Reddit and YouTube. They will collect all kinds of complaints from netizens and send them all back to the coordinator in the middle. This coordinator is the brain of the entire system and the final judge. The worker agents at the bottom who are really responsible for rolling up their sleeves and working include researchers, analysts, city builders, and even film producers. They all collaborate around the bulletin board I just mentioned and perform their duties without confusion. Of course, you also know that there are a lot of meaningless complaints on the Internet. You definitely don’t want to waste precious computing resources on solving fake problems. This is where the verification step comes in handy. The system will automatically score each captured issue based on the rating table on the screen. It will evaluate whether the frequency of this pain point is high and whether the degree of pain is great. Is there any way to solve this problem with AI? And the gap between the existing solutions in the market. And the degree of compatibility with the creator's own direction. Through this table, the system can automatically and accurately throw away those low-value bad Braille words. into the trash can, and the passing line that determines the life and death of these ideas is set at 65 points. As long as the issue is lower than 65 points out of 100 points, this assembly line will put it aside without mercy while handling their own research, analysis and construction tasks. As a result, there is no stepping on each other's toes, and there is no confusion of repeated executions. To be honest, this is the most shocking power of the shared Kanban board as a coordination, and even crazier in this single demonstration run. This system has seamlessly completed a total of 97 independent tasks, a total of 97 steps from detection and scoring analysis to the production of preliminary results. All the processes are as precise as the gears of an advanced clock. This number can be said to have completely broken the curse of the previous multi-agent system, which often collapsed after running three or four tasks. It shows a super amazing throughput. Part 4: Human Safety Net. Why 100% automation is a trap. After hearing what you have just heard, you may be thinking, it’s so cool. I can make money automatically. If you often read those farm titles on the Internet, you must think that this is very possible, but the author here gives a very strong warning. The pursuit of 100% full automation is definitely a dangerous trap. So the most critical point of the entire system is this human gate. From the initial detection to the proposal of the plan, this section is indeed all automatic. But before really starting to spend a lot of resources to construct a solution, the coordinator will send the proposal directly to the author's Telegram. At this time, as a human, you can decide whether to approve, shelve, or ask it to modify. Technically, you think you can cut this step off and make it fully automated, but please, don’t do that. The existence of this security network is to prevent those AIs that are full of enthusiasm but occasionally have severe hallucinations from using your money to build a whole meaningless mountain of garbage. In addition, the resilience of this system is not just a simple failure. A very exciting real-life episode was shared in the information. During the test, a system bug was encountered, which caused the briefing file produced by the AI to be accidentally stored in a temporary directory that will be automatically deleted. As a result, by the time it was finally delivered, the file had long since disappeared out of thin air. But the most chilling thing is that the coordinator AI actually discovered the wrong way in which the file was lost. Without any human intervention or additional instructions, it independently regenerated the entire slideshow into a permanent directory. This is really the most extreme display of healing power. Coming to the last part, pragmatic key points: Steal the skeleton. Don't be too busy with the hype. Let's take back our emotions a little and tell you the truth. What this system delivers to you is definitely not a magic product that can be solved in one click. Instead, it is a super-strong exoskeleton made of shared signboards and human gates. It gives you an extremely high-quality and extremely stable starting point. At the end of the program, I want to summarize with a question. By the way, let everyone think about it. With this set of shared signboard architecture, AI agents finally no longer fight with each other. They can tirelessly complete a lot of research, analysis, and all preparatory work for you. So here comes the question. When 97 tasks of tedious hard work can be perfectly automated and processed, you will be the most indispensable and valuable human being in this system. What kind of high-value decisions should you invest your creativity and experience in next? I hope this analysis can bring you some different inspirations. See you next time. See you next time. See you next time. See you next time. See you next time. See you next time.