Claude Code Multi-Agent Orchestration with Opus 4.6, Tmux and Agent Sandboxes

What's up, engineers? Andy Dev Dan here. We've got a couple massive releases to cover. Of course, there is the brand new Claude Opus 4.6. It's a fantastic model. What else is there to say? It's beating all the benchmarks. You've already heard, you've already seen this. This is not what I wanna focus on here. The real big idea I wanna cover with you today is multi-agent orchestration. The game on the field is changing. It's no longer about what the models allow us to do. As of Sonnet 4.5, these models can do much more than you and i give them credit for then you and i really know how to unlock the true constraint of agentic engineering now is twofold it's the tools we have available and it's you and i it is our capabilities it's our ability to prompt engineer and context engineer the outcomes we're looking for and build them into reusable systems to build them into powerful agentic layers that you and i can wield the true limitation is you and i so let's take another step at improving what we can do. Front ends, back ends, scripts, it's too simple for these models. So what I have here is eight unique applications that I had Claude Opus 4.6 create. I touched none of these, by the way, these are all one-shotted. I like to use E2B, use whatever you want. So Asian Sandbox is very powerful, but once again, this is not exactly what I wanna focus on here. We're going to use Asian Sandboxes as a playground to understand and to really dive into two key big ideas. is multi-Asian orchestration, multi-Asian observability. Once you put these two pieces together, you can do much more with your powerful cloud code OPUS 4.6 ASIAN. So first things first, we're going to crack open the terminal. If we create a new Cloud Code instance, on our multi-agent observability, you'll see that we have a new agent joining the session. And we have that session start hook captured. We have a rocket, and we've officially kicked off a new session. But before we touch our new multi-agent orchestration capabilities, we need to boot up Cloud Code in a different way. Close this, and you'll see we got that session end event captured here. And instead of opening this up in a classic terminal, we're going to use tmux. some powerful capabilities you'll see in a second the next thing we're going to do is i just want to make this super clear i'm going to type which and then cldyo so you can see this exact command that i'm running we're going to export the new cloud code experimental agent teams feature setting that to one we're enabling that feature so now we're running cloud code inside of a tmux session you can see we kicked off a brand new agent and this is where things get interesting if i type ls these are the agent sandbox directories that you saw here we're going to have our agents investigate and break down how we can set up these applications. So it's going to be our first agent team. Build a new agent team for each code base in this directory. Have an agent summarize and how to set it up. What you're going to see here is really extraordinary. You can see all of our events getting captured. Observability is really important for knowing what you can really do with your tool. You can see here, we're going to start streaming in all of our agent events and make sure that this is stuck to the bottom here. The first thing our agent does is it creates a task list. We covered this in our previous video. What happens? next is extraordinary. So we're in tmux. And so tmux has pains, our agent is now opening up brand new pains for each sub agent that it wants to run. Alright, so I'm going to go full screen here. So we can really take a look at this. And I'm going to downsize this a little bit. So you can see all eight agents that we kicked off here. Okay, so on the left, our primary agent set up the task list, created a team, and then it assigned a task to each member of the team. And you can see our status lines giving us the agent looks like we have haiku agents. You can see our context window and our agents are going to figure out how to run and summarize each code base, right? And so if we scroll back here in our agent observability, you can see a lot of work happening. Our haiku agents are running all the tools they need to accomplish the work. And we can dial into an individual agent here. If we click this, you can see this is all the tool calls for one of our agents. Looks like this is our primary agent here, right? Opus 4.6. You know, you can see here a lot of work is happening here, guys, right? 160 tool calls within just a minute time span, okay? We are scaling our compute to scale our impact, all right? If we hop back over here, you can see all the panes are gone now. What happened? All of our agents finished, right? They finished their work. So we can go ahead and get out. full screen mode and it's just our primary agent running now we can go and get rid of their swim lane here and this is a really powerful capability of multi-agent orchestration you want to spin up specialized agents that do one thing extraordinarily well right they focus on one task and then they finish so our primary agent now is just putting together a summary of the work done by the eight agents and you know take a look at the context window here we've only used 31 so that means that all of our other agents they explore eight different code bases And to be super clear here, guys, these are not just front ends. These are full stack applications, okay? Every single one of these are full stack applications. We can fully interact with these. And, you know, these are fully built out. They were one-shotted by Opus. Very powerful capabilities here. And none of this matters if you can't first trigger these features and prompt your primary agent to control these powerful workflows. And if you have no idea what's going on underneath the hood, right? This is where vibe coders fall apart. They don't actually know what's going on. on and so the engineers the builders and even the vibe coders that know what's going on underneath the hood can do so much more this whole idea that uh engineers are going to be replaced by this technology to me is absurd and it's because engineers are the best positioned to use agentic technology. So you can see here that if I hit control B and left bracket, I'm in scroll mode now. The annoying thing about Tmux is that it does change the controls a little bit. I'm in control mode. And if I just scroll up, you can see we have summaries for every single agent sandbox application that's stored in my local directory. Okay, so you can see here, nice summary of all eight and the primary agent knows how to spin them up. Let's push our multi-agent capabilities further. here's what we're going to do now so i'm going to get out of this mode i'm just going to hit escape and we're back in our primary window here let's go ahead and actually spin up new agent sandbox instances with each one of these applications all right and we'll do it in two teams we'll have a team of four uh mount the first four applications right because we have eight unique applications here if we run ls um and then i'll have a team of four doing the last set of applications okay um let's prompt engineer this properly all right we're going to start with the most important piece build a new agent team so we're triggering right these are information dense keywords that tell the agent we want a specific set of tools to execute all right and then i'll say using agent sandboxes so this is my agent sandbox skill and backslash command this is a special skill that i have to do wacky stuff like this use the backslash reboot and mount one through four in their own agent sandbox be sure to set up every agent so part of my workflow when i'm doing rapid prototyping what i like to do is just build it all in an agent sandbox like e to b as you saw here right this is an agent sandbox and then what i'll do is if i like one of the versions i'll copy it down locally and i have prompts for that of course so what we're going to do now is basically re-host these applications with a specialized agent team So we'll fire this off and you can see here right away our observability system picked up on that new user prompt submit event. And now things are going to get awesome again. Our agent is going to first run the agent sandbox skill and it's going to run the backslash command skill so that it understands what backslash reboot does. And then it's going to actually kick off the reboot. for every one of these directories and so our agent has figured out all of our tooling it ran our agent sandbox skill it ran our backslash command skill and now it's creating that task list again and so the task list is the centralized hub this is where everything gets kicked off from there we go okay so very cool now we're kicking off our agent team you can see we have that new pane so we have our primary agent kicking off the first agent and this agent is then gonna run that skill so every single agent has its own context window right so they all need to run the skill. They all need to run the setup commands. You can see it's running through the ETB setup process. And these are all Opus 4.6 agents. But you can see here all of our agents are getting kicked off again. And I'll just go ahead and, you know, downsize this a little bit and go full screen here so we can get a good picture of what's going on, right? All four Opus agents are running in parallel. They're each going to reboot the application. You know, what we saw here in the beginning, basically, we're going to recreate these agent sandbox instances with this new multi-agent orchestration tool. And so if we hop back to our agent observability system, Once again, you can see tons and tons of work happening, right? If we dial into one of these swim lanes, we can get a better understanding of the tools and the impact that every agent is creating. And, you know, really importantly here, if we search for our brand new tools, you can see we have these new task list tools. We should see some, if we scroll up here, you'll see where we have task update and we have task, right? So this is kicking off the generalized agent. And you can see here, this is the exact command that was run to kick off a sub agent sub agent that is executing right here and the great part about running this in tmux is of course the pains we can continue to zoom out a little bit so you can get a better view here every one of our agents has their own context window their own model their own session id and you can see they all have their own unique name here as well right sbx agent sandbox agent one two three and four right so this is fantastic so i'm focused in this individual window here inside of tmux we need to hit ctrl b and then left left to change our cursor position. What I want to do here is get the names of the other sandboxes that we didn't kick off. So I'm just going to ask the primary agent because it's actually not doing anything right now. The primary agent is sitting waiting for events to come back. List sandbox directories we didn't kick off. And you can see here sandbox agent number four has completed its setup. It's pinged a command back to our primary orchestrator agent. So what I'm going to do here is while this is running I'm going to go ahead and open up a new terminal and kick these off. Let's see if I got a clean paste out of that we're not going to kick this off in our flat window right we need tmux to get that visualization to get those multiple panes so i'm going to run tmux once again then i'll boot up cloud code and then i'll effectively run that exact same prompt then what i want to do is get the names of those agents that did not run so copy this and i'm doing a little bit of correction here on another screen. I'm having trouble copying this. And so I'm just going to ask my agent to do this for me. Copy the four sandbox directories to my clipboard should do a PB copy. There you go. And then I'll just paste this. So only post these four. There we go. And I'll kick that off. Now we're going to get the remaining sandboxes kicked up and I'm sure you may have noticed this, but if I go into scroll mode here and scroll to the top, I have run out of my API usage for today on my Cloud Max plan. So I'm using API billing. And yeah, this is going to burn a hole in my wallet. Drop a like, drop a comment. uh so that the youtube algorithm can can pay for some of this api usage but i'll escape and i'll go ahead and let this second agent start kicking off this workflow and so we're going to see the same thing and if we look at our agent observability we can see everything we can understand everything that our agents are doing top to bottom and if we go ahead and look over here you can see that we have one more agent finishing up its work, right? Sandbox agent number two is the last one still in progress. So this is a very powerful feature. We can now observe our agents in a more improved way just with Tmux, just by seeing these new panes open up as our primary orchestrator agent starts to set up and scale up this work. Then whenever we need to, we can just scale up and create a brand new team. So here's that second team of agents doing a whole different set of work in a brand new window, all right? that same process we go full screen minimize this a little bit so we can get a better view we have two teams of agents working and we have an observability system to trace the whole thing and so whenever we want to we just come in here it's got all four agents running it's going to mount these applications we're going to create a new e2b agent sandbox upload the app code base install get everything set up as if it's a brand new environment so we're combining several really really powerful ideas here that we've been talking about on the channel week after week we have multi-agent observer So we know what's going on. So we know how to improve and understand our systems. We have spaces to place our agents so that they can do whatever they need to accomplish their work without jeopardizing our local machine. And then of course, we have the new multi-agent orchestration capabilities coming out of Cloud Code on top of a brand new ultra powerful model that can run long duration tasks. All right. So we're talking about long threads. We're talking about big threads and we're handing off more and more work to our agents. That is the theme here. How can you prompt engineer and how can you context engineer with great powerful models to get more engineering work done? than ever with confidence. We want to be building systems of trust with our agents. Now, scaling up the model is always going to help with this, but this is not something we're really in control of. Whenever the new model ships, whatever it costs, we are just subjects to that. But what we can control is the great tooling we use alongside these three powerful capabilities. And so that's one of the big ideas I wanted to share with you here today, multi-agent orchestration, multi-agent observability. So you can dial in to anything your agents are doing and then course we have agent sandboxes a secure location to place whatever you want to have your agents do at scale all right and so we have two teams of four to be clear the agents are running on my device but the work they're doing is operating off the device they're using their local agent capabilities their local skills and then they are creating and operating inside their own agent sandboxes okay so our first team is all set up And if we go into scroll mode here, control B, left bracket, and we do some scrolling, we should be able to see everything set up live. So let's go ahead and open these up. I'll take these existing windows, these existing sandboxes. I'll just move them up onto my monitor here. And then we should see them open up in this browser window here. We'll take a look at the brand new tools that allow and enable these workflows in just a second. Let's go ahead and just get these opened up so we can see how our agents have done. I'll say open in Chrome. You can see these sandbox. are going to be alive for 12 hours. All right. And it looks like they did open up in this other window. Super annoying. That's fine. I will copy these four newest ones, drag them in. Here we go. So here's our agentic support. It looks like we're missing some data here and it looks like we're missing data here. Let's see if we got our gallery. Nice. Okay. So we did get some nice information there and we have our data here. All right. So very cool. And so we can continue to prompt to resolve these issues. I'm going to go ahead and just give this a shot and I'll say data is missing from This and... go ahead and go here as well basically our sandboxes were rebooted but it didn't reboot with the exact same data or with any data for those two applications so we're going to go ahead and have these agents do this work and what i'll do here is i'm going to stop this because the primary agent started working right you saw that the primary agent is trying to take over i'll say spin up a new agent team to do this work for you give them all the context they need skills sandbox backslash command and so i'm just being super verbose there with my prompt engineering with that agent can see here our second team finished so let's see how this team did yeah open all four urls in chrome okay so we'll kick these off all right so we have these opened up let's go ahead and get these drag and drop over here okay we have our mission briefings dashboard we have our portfolio application so we can track our forecast for our portfolio we have a recipe app and we have a ad dashboard to see how our ads are performing nice so we got all the data for these so all these four worked love to see that. And the two issues we had with these code bases are getting resolved here with our two agent agent team. So this is an iterative process as well. We're going to want to be firing off ad hoc agent teams to perform specific sets of work. In our previous video where we talked about the task list feature, let me go ahead and see if I have that diagram pulled up here. Yeah. So in our previous video, we talked about the cloud code task system where you prompt your primary agent and your primary agent creates a task list that multiple agents, basically an agent team operates on. This is a very powerful feature that is taken to the next step with the multi-agent orchestration feature that you can now tap into. But the idea here is really, really important. As you're building out real features, as you're scaling up the work you can do with your agents, it's not just about a single agent or even a couple agents or even parallel agents. You want to be building teams that communicate together that are all driven toward accomplishing one specific goal. Think about building out a feature. That's way too much work, especially as you enter real legitimate production code bases building out a full feature requires organization it requires communication right and so this new agent orchestration feature allows us to really tap into that brand new system that way of thinking all right and this is what this feature looks like end to end we're going to create a team we're going to uh create tasks which spawn agents they all work in parallel and then they shut down and then we delete the team and we'll look at the tools here in just a second but a really important aspect of this is that when the work is done you want to delete the agents this forces a good pattern of context engineering where you reset the context and start over so you can see here's the agent shutdown process happening from our primary agent these are shutting down the tasks are gone the pains are now gone and so apparently this has fixed the issue you can see here db was intact we restarted uh both processes were down so we should be good on these two uh applications let's refresh still not good there that's too bad and let's refresh here it looks like this one did load its data here let me see if that actually worked loading pull requests and still nothing here all right so we have issues here i don't really care about these you get the point here right we got six out of eight sandbox environments spun up in brand new systems and you know just to be super clear and transparent here are the new environments and the new urls here are the old ones i just want to show that i have these and that these are actually different urls all right so these are all unique asian sandboxes and we can be even clear about this i have a bunch of these running right now if i go to a terminal here and I say, I pass in this skill. We should spin up a new agent for this. We always want to be operating on fresh instances. We don't need Tmux for this. This is going to be a single clock code instance. And I just want to show you all the current running agent sandboxes. So I'll kick this off. I'll fire off my agent sandbox skill and I'll just say list all running sandboxes. And so we should see, I don't know, some 20 or 30 sandboxes here. You can see it's validating my E2B key. And then let's go ahead and see. There we go. Working through some issues. And of course, as a sandbox list, you can see we have 24 four sandboxes. Let's go ahead and get that list just to make it super clear here how much compute we're running in parallel. There we go. Looks like that was the command. It's getting the information for each sandbox environment. I'll link a previous video where we talked about how you can set this up and how this really works in the description for you if you're interested. I think agent sandboxes are and will be a big, big, big trend as we scale up what our agents can do on our behalf. You're seeing this with the whole Mac mini craze. As you can see in the background, I have a Mac mini. I've had this thing for a while and was, you know. decent amount ahead of that trend. I'm having my agents run in its own dedicated environment. As you can imagine, I have multiple sandbox environments, one like E2B operating purely in the cloud. And then for more personal workflow where privacy is important, you can have agents run. course on your own local devices like everyone is doing with multbot or clawbot or whatever it's called now but you can see here you know i have 24 agent sandboxes running and you know you can see i have duplicates of a lot of these things that we've been looking at here right so i have multiple versions of this just want to make this super clear i have a skill i have an agent that operates and can manage all of these agent sandboxes at scale this is going to be really important moving forward when you're scaling your compute to scale your impact which is the big theme of everything we're looking at right now all right here we look at how to spin up teams of agents okay and it all comes back to things we talk about on the channel all the time all these fantastic new tools coming out the cloud code team all these new capabilities there is a lot of engineering work they put into this you know big shout out to the cloud code team but i do want to say that underneath all of it is a concept we always discuss it's the core four context model prompt tools everything boils down to that all right everything is the core four okay and just quickly you know we saw all this work happen we saw our multi-agent system track all of this and you can see all these new task tools team delete team create we have these new task tools task create task get so what are all the new available tools let's go and take a look at this we have kind of three categories of tools that this you know new multi-agent orchestration system gives us team management task management and communications Team create, task, team delete. Task has been around for a long time. This is how you kick off an agent in parallel. But then we have all the new task management tools, right? Task create, task list, task get, task update. But the most important one of all probably is this send message. This is how the agents were communicating. And after they communicate, after they do all the work, as we described here in this workflow, right? This is kind of the multi-agent orchestration workflow built up with this new tool. It's this, right? Create the team. create the tasks, spawn the agents, work in parallel, shut them all down, delete the team. This is the full workflow of the brand new Claude multi-agent capabilities. So with every new capability, with every new feature coming out of Claude code, coming out of all these agentic coding tools, with every new release of the new model, the question is always the same for you and I, the engineer with our boots on the ground, working with this technology every single day. How can we understand the capabilities available to us to accelerate our engineering work? improve, tools will change, and that means that you and I will always be the limitation it's about what you and i can do so with every feature release make sure you're digging in make sure you're understanding what's available to you so that you know what you can do every engineer is limited by their tools and their knowledge of their tool so that's why multi-agent observability is super key throughout any point in this workflow we can jump in here and we can investigate and see the communication see the tasks between our agents that we kicked off we can see all the events i'm going to leave my multi-agent observability updated to support all these new tools link in the description for you and i'll also link a previous video and my agent sandbox skill for you to play with again link in the description for you if you're interested in taking your agentic coding to the next level check out tactical agentic coding this is my take on how you can accelerate far past ai coding and vibe coding with advanced agentic engineering so powerful your code base runs itself we're seeing multi-agent orchestration come out of the cloud code team we have had this documented and covered inside of this course inside of agentic horizon specifically we have had working versions of multi-agent orchestration for months now this is all here you know a lot of the ideas we talk about on the channel are taking to the next level inside this course if you're interested check this out i know a lot of engineers the channel have already checked out this course and you know to be super clear there are thousands of engineers that have taken this and then have gotten great value out of this so i'll leave this linked in the description if you're interested as well we're going to be covering multi-agent orchestration a lot as we move forward it's going to be a big trend because it allows us to do what we always do on this channel scale our compute to scale our impact thanks so much for sticking around you know where to find me every single monday stay focused and keep building

🔥 The ENGINEERING game on the field is CHANGING. It's no longer about what the models can do — it's about YOU and your ability to orchestrate them.

Multi-Agent Orchestration is HERE and if you're not scaling your compute, you're leaving engineering power on the table.

🎥 VIDEO REFERENCES
• Opus 4.6: https://www.anthropic.com/news/claude-opus-4-6
• Claude Code Orchestration: https://code.claude.com/docs/en/agent-teams#set-up-agent-teams
• Multi-Agent Observability: https://github.com/disler/claude-code-hooks-multi-agent-observability
• Agent Sandbox Skill: https://github.com/disler/agent-sandbox-skill
• Tmux: https://github.com/tmux/tmux/wiki/Getting-Started
• E2B: https://e2b.dev/

PUSH YOUR AGENTIC CODING BEYOND
Tactical Agentic Coding: https://agenticengineer.com/tactical-agentic-coding?y=RpUTF_U4kiw

🚀 In this video, IndyDevDan breaks down the brand-new Claude Code multi-agent orchestration capabilities powered by Opus 4.6. We go beyond the benchmarks and dive straight into what matters: how to spin up Claude Teams of specialized agents that work in parallel, each with their own context window, their own model, and their own tasks. Watch as we orchestrate two full teams of four Opus agents running simultaneously across eight full-stack applications using Tmux, E2B agent sandboxes, and the new Claude Code experimental agent teams feature.

🛠️ This is tactical agentic coding at its finest. We demonstrate how to prompt engineer your primary agent to create task lists, spawn sub-agents, assign work, and coordinate results — all within a single orchestration workflow. From team creation to task assignment to parallel execution to shutdown, you'll see the complete Claude Code Orchestration lifecycle in action. The new tools — TeamCreate, TeamDelete, TaskCreate, TaskList, TaskGet, TaskUpdate, and SendMessage — give you full control over agent orchestration like never before.

🔍 But orchestration without visibility is chaos. That's why we pair multi-agent orchestration with multi-agent observability. Using our open-source observability system, you can trace every tool call, every agent session, and every task update across your entire agent fleet. We tracked 160+ tool calls in under a minute and maintained full visibility into what each agent was doing. This is how engineers build systems of trust with their agents.

💡 The core lesson: everything comes back to the core four — context, model, prompt, tools. Opus 4.6 gives you the model. Claude Code Orchestration gives you the tools. Agent sandboxes give you the space. But the true constraint is YOU — your ability to context engineer and prompt engineer the outcomes you need. Vibe coders fall apart here. Engineers thrive.

🌟 Whether you're scaling agentic coding workflows, building agent observability pipelines, or exploring agent orchestration patterns, this video gives you the playbook. Scale your compute to scale your impact.

Stay focused and Keep Building.

#agenticengineering #claudecode #agenticcoding