SEE CMUX SOLVE Multi-Agent Orchestration (Claude Code and Pi Agent)

As engineers, there's just one idea you and I need to win in the age of agents. Scale your compute to scale your impact. The goal is clear, but the path is less clear. The big open question the best engineers are asking is, how exactly should you orchestrate your agents for maximum impact and there's no shortage of options boris and peter are saying loop engineering is the future and prompt engineering is already dead i completely disagree with this there's the ralph style to-do list to keep your agents focus on a massive list of work fable is going to love ralph then there's classical sub agent delegation and many many more patterns i've been engineering for over 15 years and building with agents since it was first possible and we have to call out the fact that these ai labs have massive incentives to keep you and i spending tokens token maxing when the truth is there's a dozen different agentic patterns you can use to ship with agents One of my favorite patterns is three-tier agent orchestration. Orchestrators prompt the leads. Leads prompt the specialized agent experts. But every multi-agent approach has similar problems. The big three being, how do you start, interact, and improve your growing fleet of agents? If you're vibe coding or token maxing, This doesn't matter at all. You can spin up 20 agents in a loop and look the other way. On this channel, we don't vibe code. We agentic engineer because an agent you can't monitor is an agent you can't improve. I have three multi-agent orchestration problems I'm going to solve and I'm looking at CMUX as a potential solution to this problem. As we work through the feature set of CMUX, I want to share with you how I'm continuing to learn new tools and technology with agents. If these ideas interest you, smash the like button, lock in and let's improve our agent orchestration. Whenever I sit down to learn a new tool or technology, I have a state-of-the-art agent build out a comprehensive HTML file that I can use to visually understand, study, and master new tools. So that's how we're going to learn CMUX right now. Let's start with the problem. What problems are we going to solve? I don't start with tools. Problems come first, tools come second. I know I'm late to the CMUX game, but I don't really care. Outside of experimentation, if I don't need a tool, I don't pick it up until I have a problem first. I have three problems surrounding multi-agent orchestration that you likely have too. Let's see if CMUX can solve three. popular multi-agent orchestration problem so first off no programmatic access to your agents this is a massive problem because it means you're the bottleneck and it keeps you in the loop all the time meaning your agents can never move at the agentic speed the big question is Can CMUX solve this? The answer is, of course, yes. The core value proposition of tools like CMUX and the tried and true TMUX is this exactly. They give you agentic access to every single terminal. An agent you can't see is an agent you can't improve. This is going to be increasingly controversial. A lot of engineers are turning off their brain. They're not studying what their agents are doing. They're spinning up fleets and they're ignoring the journey, all the mistakes, all the successes, all the things that they would want to emphasize, reproduce. and cut out of the workflow is completely ignored this is one of the key differences between vibe coding and agentic engineering an agent you can't see as an agent you can't improve i need to be able to see every one of my agents okay it doesn't matter if it's cloud code pi open code codex i need to be able to see it monitor it and understand when things are going right so i can re-emphasize that behavior in the agent via the core four or i need to know when things are going wrong so i can present negative rewards okay so can see solve that problem for us yes it can we have per workspace color role icons identity we have tabs we have banners we have everything we need to quickly jump into an agent and see it. Quick teaser here. I have CMUX ready to go. And you can see here, we have a great breakdown, the kind of fundamental pieces of a tool where we can quickly see, observe, and therefore improve everything, right? We have a Cloud Code agent. We have a Codex agent. We have two PI agents running the Minimax M3 and the JLM 5.2. We'll get back to this in a second. Let's fully understand every single problem. And this is a big one. It's increasingly becoming a problem as I deploy more and more agents and set up new teams to do specific sets of work extraordinarily well. If you're building custom agents, you're running into this problem too. If you're deploying five, 10, 20 plus agents, you have also seen this problem. Booting an agent team by hand kills the agentic speed. It just takes too much time. So can Tmuk solve this? Of course, yes, it can. It has reusable session files, but more importantly, this tool has agentic access. If a tool has agentic access, that means you can move at the agentic speed and solve. problems in natural language with well-constructed prompts. Okay, so these are the problems we're going to solve. CMUX has solutions to all of them. I think the question is, do you have to use CMUX or is something like TMUX going to get you all the way there as well? So let's dial back into CMUX and understand what this tool can do and see if it can help us solve these three key problems. Agentic access, monitor to improve, and quick agentic launch for your thousandth agent run, not your first, not your tenth, not your hundredth. So what is our agent teaching us about CMUX? Let's understand the mental model that really builds the foundation of understanding CMUX and other terminal multiplexes as well. You have the window, which is going to be this entire window here with Control-Shift-N. We can boot up a brand new window. And then we have our workspace. Workspaces, you know, we can boot up with Command-N. And this is going to give us a new workspace to set up teams of agents, to set up groups of terminals, to do whatever you need to do. Then we have panes. The pane is going to be the actual window. We have a pane. here we have a pane here we have one here here and here it's important to mention if we do command t this opens a new surface inside of this existing pane so we have individual windows You can separate and group your agent teams, your terminals. You can have production. You can have staging. You can have dev. We can have workspaces. So for me, I'm going to put my agent teams on the workspace level. And then inside of the workspace, you have your individual pane. So we have a pane here. In my multi-agent orchestration workflow, I'm going to put my lead here. And then the workers go over on the right. So I can look at the level that I need to to get the job done while not losing the depth of visibility into the low level worker agents. This is super key, right? Prompting in a black box in sub agents is a great place to start, terrible place to finish, right? We can do a lot better than that with the right agentic tools. That's why we're trying CMUX to see if we can really improve our multi-agent orchestration abilities. Okay, so that's the mental model of CMUX. And so how do we control it? It's pretty simple. Thanks to the programmatic access, we do four things. We can send information like send key. We can then read the screen after the action has been completed. And then we can open and close surfaces and the loop repeats. Okay, so this is how you control it. We're going to walk through concrete examples of how you can use CMUX in just a moment here. But here's an example of the API, right? Send, read. close surface you specify the surface and the exact text you want to send down another powerful part about cmux we can build an orchestration agent around pushing and pulling specific events from the cmux lifecycle and therefore our agents lifecycle okay so there's a lot of customizability here when i sit down to learn a tool i use agents the entire time learning is the killer use case for agents coding comes next in these html rich guides i want an incremental tier based explanation of how i can best use these tools to solve my specific use case and that's exactly what my agent has done here a lot of really smart engineers are turning their brains off i want to push back against that we want to rise with the ceiling of agentic engineering not the floor of vibe coding okay so um let's run this right let's understand cmux at a deeper level see here this tool can get very very complex it can help you orchestrate and do a lot of work they even have like a mobile app here you can access browsers there's a lot of cool stuff you can do with this tool we're not going to cover all that we're going to focus on the problems we need to have solved whenever i'm learning a new tool i like to start super simple understand the foundations of the tool, and then move to more and more complex scenarios. All right, so let's go ahead and start from scratch here. I'm going to open up a terminal, boot up a cloud code opus window here. If you're working in the future, you have access to Fable. I am jealous of you right now. Can't wait for that model to be turned back online, although it looks like there's going to be massive delays with getting that back online. So I have a CMUX skill that was built that just makes the agent aware of the tool. Open a workspace in a fresh temp directory, create 60 throwaway files, LSLA, read the output back, and then continue from there, all right? So we're just having our orchestrator agent operate on CMUX windows. So let's see how this looks. And I'll hold them both open here at the same time. So you can see here our existing agent team window is open. My agent is operating this window, okay? So we have a new workspace here inside the window. We have a pane with a single tab. You can see there's that LS. You know, I'm not typing anything here. This is my agent operating a workspace, to be clear. This is bigger than a terminal. It's done, right? So we had it LS. all these mock files are created right if we close this door ls ourselves we can see them exactly the great part about tools like cmux and tmux is that you can jump in whenever you need to write the agents stopping black boxes when you can actually touch and jump in the terminal window agentic access here is super super important very simple great place to start when you're learning a new tool just show a basic use case of the app so let's split into a grid i want to see a grid here so i'm going to close our throwaway 60 and we'll work our way up to you know a full multi-agent orchestration example here we're just setting the groundwork always start with the fundamentals so this is going to open a two by two grid in a new workspace There we go. And in each workspace, it's going to run specific commands. On the right here, it's just an empty terminal. Here we have the date being updated per second. And then we have a tick in the bottom left. And then in the top left, we have htoprunning. Each one of these clearly documented in the workspace name. You can see how this can be very, very powerful for agents, right? So not only is this tool helping us move at the agentic speed, it's giving us access to quick start commands to run repeat workflows. All right, so. boot up your front end, boot up your back end, boot up all your services, boot up your database. Let's keep pushing to see how this tool is differentiated from a tool like Tmux because Tmux is the OG of this. Let's go and tidy up. Let's have our agent clean everything up. Take the dash workspace, rename it. flash it and then start closing it okay so this is where we start pushing into the value prop of cmux it can control the panes and the windows in a very very precise way there's the flash it updated the name to retired so we have live updates happening at any point in time that we can agentically program we can prompt engineer into the system to really control the experience of agentic engineering and that's really where this all goes right the ultimate user interface for agentic engineering, the DX for agentic engineering, it is not set in stone yet. As much as everyone's talking about loop engineering and building these new DXs, there are no winners yet, okay? Only experimenters, only researchers, only those pushing what they can do with agents. So keep looking for the optimal developer experience for your work, for your agent, so on and so forth, right? And that's what we're doing here. We're just understanding the state space of the available agentic tools. okay and as you'll notice here building blocks extensibility adaptability plugability is a key part of great agentic technology, right? Which is why I'm drawn to CMUX in the first place. Great, agentic access, rename, flash, close the whole workspace. Great, you get the picture. Let's get more complex here. Let's boot up Cloud Code and the Pi Coding Agent side by side. We're gonna do this in a brand new workspace. That's the mode I like to work in here with CMUX so far. I'm getting into this nice pattern of creating new workspaces, not windows. Two columns here side by side, and one is Pi and one is Cloud. And so another great piece about this, you can orchestrate any agentic coding tool you want to. If you're a cloud code guy, scale up your cloud codes. If you're a Codex fan, scale those up. If you're an open code scrub, go for that. Just kidding, just kidding. If you're a Pi coding agent, elite agentic engineer, plugging and playing with your own custom agent harnesses, then you'll go for that, right? Opus 4.8, one of the best orchestrator models in the game until Fable supersedes it. You can see here, we are orchestrating multiple coding agents with the orchestrator agent. Okay, very, very powerful stuff. I'm using IPy for my Pi coding agent. This is my customized. PI coding agent with several different extensions built into it. There's a specialized developer experience as an agent harness. So we've just asked a simple question here about processes versus threads, and they both answered, right? So nothing special there, but the value proposition is obvious. You can boot up agents and you can prompt agents using CMUX, and we can continue to scale this, okay? So let's do exactly that. Let's start orchestrating fleets of agents, okay? So let's do a 2x2 fleet. And a really important piece here that you'll also notice is that These agents have completed. They then send a notification to this top window. Okay, so say you're working over here, getting things done on another branch of work, another set of work. CMUX is going to notify you when this set of agents or single agent, if you want to do one workspace per agent, it'll let you know when that work is done. Okay, but here we go. We have a new fleet. This is our security fleet. Send up four coding agents, Cloud Code, Codex, Pi with Minimax and GLM. Create a fleet. List the top three security vulnerabilities you can find in this repository. So this is a simple mock repository. This is a waste for these agents. They're not really going to do anything. But you can see here, this is a simple. important idea of literally scaling your computer to scale your impact say you were looking for security vulnerabilities in your code base you can spin up one agent and have sub-agents under it run or you could set up different agentic coding tools with their own unique advantages running different agents to run the exact same validation pattern the exact same security checks across your code base okay this is the simplest clearest example of scaling your compute to scale your impact the big difference here with a tool like cmux or tmux is that this is fully visible okay so we have minimax m3 giving a shot at this glm 5.2 and of course the state of the art models gpt 5.5 and opus 4.8 they're all applying their opinions their perspective their expertise their specific model advantage against this problem of checking for security vulnerabilities a big topic right now in the age of agents one of the key pieces of this is now inside of our orchestrator because remember the way i'm looking at this tool is any point to be able to use my orchestrator to drive results in any one of my windows any one of my workspaces and then the individual panes but also have the ability to go into any individual agents and really just prompt out and dial into any details we need to on an agent level okay so this is important we want to be able to jump into the process at any point understand what every agent is doing who's got the advantage who's running patterns we want to replicate and who's doing stupid we don't want to do again okay and that's on a agent decoding tool level all the way down to of course the model level and then further on down right context model prompt tool system prompt context so great these are all done and the orchestrator is reading it back out this is now in idle mode because the orchestrator has read the response it's read the surface right it knows that things are done this is where we start getting out of what tmux can do right we're working on multiple problems we're scaling our computer scalar impact and we can see everything and we can understand everything okay Fantastic. So we are doing multi-agent delegation. But of course, one of the key things here is any agent can prompt any other agent. So I'll just try to do something random here. Inside of this cloud code worker, I'll type CMOX to activate the skill, ping your other fleet coworkers. So very powerful stuff here. It's going to read the screen of each surface. I've been seeing really, really great success with this three-tier architecture with my agents, right? You have a top-level orchestrator. You then have team leads, and then the leads have workers. that actually do the work. Okay, so we're replicating traditional hierarchy structures without the true cons and true problems with hierarchies, right? This is not a top down structure. Any agent can prompt any agent, you can see that here, thanks to this skill here, we're getting a nice simple prompt. coming into every single agent from inside cmux right so our agents can prompt our agents this is a flat hierarchy even though there is one orchestrators at the top and so i like to put my team lead on the left as you see in a moment but we can just drag and drop this down here or we'll move it over here and this very quickly kind of shows you the organizational structure that you can build lead here workers here and you can build any structure you want the point is is that the communication channel is flat right so fantastic what else can we do with this tool let's scale into workspace Let's really push what we can do with this. So I'm going to copy this larger prompt here. I'm going to have my agent do this. Here we're going to showcase like the importance of organization inside the workspace. You can see here we're starting to already have a lot more organization than normal if we were just in a single terminal or if we were using a single terminal with tabs. We're getting notification events which are pretty pluggable. You can do a lot more with them than we are. You can see here imagine you're booting up multiple teams to solve multiple problems. We have gamma, beta, alpha. We can jump into these and we can see our let me pull up the orchestrator so we can see what going on here right our orchestrator getting work done booting up sets of agents programmatically let's close our security fleet here okay so we have gamma we have beta and we have alpha booted up okay so you know imagine any multi-agent orchestration set of teams you want that's a key value proposition here You have multi-agent coordination at a massive scale, right? Just an absolutely massive scale and you're in control and you have visibility into it. Okay, super, super important. If you hold command, you can see all the hotkeys for each workspace. This is clearly the unit that CMUX sees as the most important to be navigating. It's at that workspace level. But of course, we can quickly move through all of our surfaces as well, right? So again, just expressing the idea here, you can scale agents across multiple workspaces as well. So let's look at the scenario where you need to fix a bug right away. Say production's going down, your customers and your users are completely blocked. So you're losing money by the second. Okay, so I'm gonna hit new here just to reboot everything and then paste this in here. There's a longer prompt, but we're gonna have agents race toward a solution. Okay, so I'm gonna close all these windows here. Say there's a production issue. and you need to resolve it yesterday. This is going to be a differentiating agentic engineering pattern that you can use. It's super simple. Throw more compute at the problem. Have your agents race toward a solution. We're going to do a needle in a haystack search. In reality, this will be a hot fix you'll need to deploy, but you can just throw a bunch of agents at a problem at the same time in parallel. You want the first available answer so you can deploy the fix back into production. Okay. So in this scenario, we're going to boot up an eight agent race to just race toward a solution. So there we go, all the windows popping open here. And we're going to deploy a variety of compute to get the job done. This is needle in a haystack, capture the flag, first agent to the goalpost wins type of task. And multi-agent orchestration lets you do this really, really well. Every context model prompt, every agent decoding tool is going to have their unique advantages, strengths, and weaknesses. And so when you need to deploy something quickly, when you have a hotfix you need to deploy, growing. different types of intelligence at the problem is going to get you the better result pretty much every single time if you're willing to of course pay for the compute and so you can imagine you know you have your red team agent race here and then you might have another set of agents in a you know your infrastructure agents or your deploy agents here and then you might have your review set of agents here to at scale quickly review the issue so that you can then hand it off to your infra team of agents right your deploy agents and now we have a bunch of compute thrown at the problem. We've got Opus, we've got Sonnet. And another advantage here, you can see a couple of my PI agents didn't even fire. It looks like an environment variable setup problem. That's fine. My local agent, my local Quinn is working. All the codecs and all the Sonnet models quickly found the solution. That's done. And now this issue can be reported. And now we're waiting for that notification event for my orchestrator and needs input. There we go. Okay, so it finally triggered there. And now our orchestrator should pick this up. Let's go and see if it does pick it up. because all the agents are complete. Maybe this is a good time for us to transition a little bit and talk about the pros and cons of this tool, right? And maybe my agents just don't understand the tool properly. That's fine. It could be a mistake on my part. When you're using tools like this, these new tools, tools that are just kind of appearing overnight. And I know CMUX deserves more credit than that. This is a solid tool that's been around for a while. There's a whole wave of agentic tools coming, and you're always going to work and compare that against tried and true technology. So when I sit down to solve problems and learn with agents, I'm always comparing against what else exists. For a lot of the capability we've looked at here, Tmux can do the job. The thing is that it's not going to be as customizable and aesthetic. And the developer experience, I think, with Tmux is a little harder to get right. That's what Tmux does really, really well. It's made for the modern era of building with agents programmatically on Mac devices. Okay, so yeah, probably should have mentioned that earlier. If you're on Linux, Windows, WSL, you're going to need to use... a tool like Tmux, this is Mac only. The big problem with this is just maturity, right? These brand new tools have issues and it could be me, but you can see here, you know, in our red team response here, our Orbitrator has done nothing. Check and understand why you stalled there, right? Because really we wanted to ping on the first completion of our tool, right? Our first completion of any window. Okay, so foreground 10 minute. Okay, very weird. So we did get a notification event. It just seems like our agent didn't register it. This is one of the things I need to spend more time on to make sure I'm getting right. And it's just one of the things that you'll have to understand when you're building and teaching your agents how to use the tool. Something to be improved there, but the high level is still there. There are notification events you can plug your agents into to wait for responses from your agents. We're going to stand up eight agents across four workspaces, two agents in each, give them each a two-column layout. The left is the agent, and the right is going to be a browser. Now, this direction is fine. CMUX is trying to grow as a product, so it has in-app browsers. That's really cool. The real killer feature here is programmability. This is the most important thing. It's nice that it's lightweight. Of course, open source is super important, but customizability is what comes next, and there's a whole slew of customization you can build out in. to your cmux application that we're going to gloss over because these are extra features you can change the theme and feel you can add custom action buttons that's all well and good the most important thing here by far is agentic access right this is a huge theme this is one of the key pillars of agentic engineering right it's making sure that you have agentic access to every tool to every service every product that you want to interact with at the speed of agents not the speed of humans i think we can just say that right in the digital world eventually the physical world too we move very slowly now and there are advantages to moving slowly coding is not our domain anymore that is the domain of agents Okay, but orchestrating intelligence is our domain. This is one of the five pillars of agentic engineering. We covered this in a previous video. I'll link that below if you're interested. Here's the kind of high level of the five pillars I'm really, really focusing on right now over 2026. And agentic access is number five. Even your agents should have agentic access, right? You want that meta level of control. And this is all to raise the ceiling of your agentic engineering, not the floor. of live coding when these missiles class models come out a lot of engineers are just going to sit they're going to get lazy they're going to stop thinking and just hand the wheel to the agent on the channel we're not going to be doing that if you agree with that ideology like follow subscribe all that good stuff the kind of one-liner is if a tool does not have programmatic access i just completely ignore it now i don't give it another second of my time okay agentic access gives you the agentic speed let's see where our team is fs delta gamma beta alpha team imagine these are you know the names of features the names of bug fixes the names of things you are actually building against we have a nice browser on the side and we have a two agent team per workspace to work on that browser okay so you get the idea here right we have a little custom theming here to help us get back in the loop when we need to be in the loop to orchestrate our agents So again, you know, having visual identifiers improves your developer experience, lets you focus on what matters the most, and then you can just move on. The biggest risk with CMUX is that this is a new, relatively fast moving tool. I'm going to be using both of these tools side by side, not so much using warp anymore, but CMUX and TMUX I'm going to be spending some time on. And that leads us to the kind of final problem. I need to be able to quickly launch agents for the thousandth time. So let me show you really concretely how I'm using a tool like this. I have a just file inside of my code bases, and this lets me give a really quick one tap command to boot up new agent teams. Okay, so I'll just type J fast CC. If I hit enter here, I need to pass in a feature. So what's the focus of this agent team? Okay, so I'll type fast CC, and then we'll call this SSC. Let's assume we're adding server sent events into our web server. Okay, so we'll hit enter there. And this is going to pretty much instantly. boot up an agent team and this is a unique asian team in the multi-agent orchestration pattern that i like this is three tier multi-agent orchestration and it looks like this the orchestrator here team leads here every one of these is a team lead in my workspace and so this specifically is the team lead right now i'm using a glm 5.2 and then every team lead gets access to worker agents okay so here's my plan agent here's my theoretical build agent here's the build front end and here's the testing agent might seem complex to add one additional level of agent but it helps a lot and i've needed a tool like cmux tmux wasn't cutting it the right organizational structure and so i'm picking up a tool like cmux to kind of help me solve this multi-agent orchestration problem these sets of problems it's starting agents it's monitoring agents so you can improve them and then it's being able to scale your orchestration to whatever level you need to and so as mentioned this is not top-down agent communication right any agent can prompt any agent i also have my specialized commsnet pi agent coding extension but cmux lets me push outside of the extension and have any agent decoding tool talk to any agent decoding tool thanks to the cmux skill bi-directional flat agent communication is super important while maintaining an orchestrator a lead and then a worker level setup and i'll link the video where we go into multi-tiered asian orchestration this isn't for everyone a lot of engineers are going to look at this and say why do i need this it's slop this isn't useful that's fine it just means it's not for you okay as i mentioned in the beginning there's many ways to do this you can use the ralph to-do list you can use the new loop engineering patterns and you can just use classic cloud code sub-agent delegation that's fine right they have the slash workspaces they have slash loop they have slash goal right use whatever works for you my goal here is just to show you the optionality of what's available to you and to re-emphasize agentic access and full control over the primitives of agentic engineering right context model prompt tool once you have an agent up and running you can scale it right now i have a team here that has let's say these are each 1 million contacts each specialized that gives me 5 million tokens of context to work with quite literally right that's not an exaggeration when you put together your agents in a way where they can all communicate you're getting a bigger team of intelligence to work on your behalf okay so um you know what's the verdict i like cmox i'm going to be using this tool it stands up to tmux pretty well i'm going to be keeping my eye on any glitches or bugs with this tool. But so far, I'm a big fan. It's the Scriptable Service API, the agentic access that really matters with this tool. There are a bunch of other features here I didn't mention. I'm sure some CMUX hardcore users are going to be flaming me for this being so relatively simple. But the key here is it's programmable, right? That's the most important thing by far. Agentic access is a requirement. for agentic engineering and this is going to compound every single thing you set up here from this five item list is going to give you a compounding advantage for agentic engineering once again we want to be moving up with the ceiling of agentic engineering not the floor of vibe coding all these ideas are big ideas we talk about and really break down in tactical agentic coding. I'll leave a link in the description for that if you're interested in paying a little bit to get a massive advantage. If you're not though, that's totally fine. I'm going to leave my entire guide code base here to really break down CMUX capabilities. If you made it to the end and you want to see more agentic engineering content like this, drop a like, share the video with a friend, with a coworker, and then comment down below, how are you thinking about multi-agent orchestration? Do you have any unique experiments or tools you're running to better scale your compute to scale your impact. You know where to find me every single Monday. Stay focused and keep building.

🔥 Hot take: An agent you can't SEE is an agent you can't improve. 

Spinning up 20 agents in a loop and looking the other way isn't agentic engineering, it's  gambling with tokens, its vibe coding... 

🎥 VIDEO REFERENCES
- Learn Cmux Codebase: https://github.com/disler/learning-cmux-with-agents
- Cmux: https://cmux.com/
- Five Pillars of Agentic Engineering: https://youtu.be/2KcITKKJikA
- Pi Agent Teams Video: https://youtu.be/M30gp1315Y4

🚀 BECOME AN AGENTIC ENGINEER
https://agenticengineer.com/tactical-agentic-coding?y=WAFUMBLOjHo

The AI labs WANT you token-maxing. I'd rather give every one of my agents agentic access and watch Cmux solve the three multi-agent orchestration problems that are quietly killing your agentic speed.

🚀 In this video, I (IndyDevDan) learns Cmux live: the way I learn every new tool, with AI coding agents driving the whole session. Problems first, tools second. We line up three multi-agent orchestration problems every engineer building agent teams runs into: no programmatic access, no way to monitor and improve, and the slow grind of booting an agent team by hand: and we test whether Cmux (and the tried-and-true tmux) can actually solve them. The answer comes down to one word: agentic access.

🔥 Watch as we scale from a single throwaway workspace up to an 8-agent hot fix team racing toward a solution. Claude Code, codex, and Pi agent running GLM-5.2 and minimax-m3 all attack the same problem in parallel: first agent to the goal post wins, then notifies the orchestrator. This is scaling your compute to scale your impact, fully visible, no black boxes. We boot 8 agents across 4 teams with 4 in-app browsers, and any agent can prompt any agent:  a flat communication channel that still keeps an orchestrator, leads, and workers in their lanes.

🛠️ The pattern I keep coming back to is three-tier agent orchestration: orchestrators prompt the leads, leads prompt the specialized agent experts. Cmux controls panes, windows, and surfaces with four simple moves: send, read, open, close. So your agents can drive terminals at the speed of agents, not the speed of humans. I show the exact just-file one-tap command I use to instant-launch a fresh agent team for the thousandth time, because booting agents by hand kills the agentic speed.

💡 The core lesson: if a tool doesn't have agentic access, I don't give it another second of my time. Agentic access is one of the five pillars of agentic engineering — and it's why I reach for Cmux over a single terminal or even sub-agents prompting in a box. Monitoring isn't optional. An agent you can't see is an agent you can't improve. That's the line between vibe coding and agentic engineering, and it's the whole reason we rise with the ceiling instead of sinking to the floor.

🌟 Whether you're orchestrating Claude Code fleets, customizing your Pi agent harness, or designing your own multi-agent orchestration patterns with agent teams of AI coding agents, this video hands you the mental model, the control loop, and the workflow to learn with AI agents and command your fleet. Stop token-maxing. Start orchestrating.

Stay focused and keep building.
Dan

📖 CHAPTERS
00:00 Multi-Agent Orchestration Patterns
02:02 Agent Orchestration Problems
05:40 Cmux in 3 Minutes
07:14 How Your Agents Control Cmux
10:17 Cmux Grids, Terminals, and Agent Fleets
18:53 8 Agent Hot Fix Team: Race and Notify
23:01 8 Agents, 4 Teams, 4 Browsers
26:04 How I'm Using Cmux for Instant Sessions

#agenticengineering #aicoding #terminal