Anthropic buried the lead because while everyone's been freaking out about Opus 4. 6 benchmarks, they quietly released a new experimental feature inside of Cloud Code, agent teams. And after spending the last few days testing it, I believe they've laid the foundation for something really exciting.
And that's because agent teams aren't just your typical sub agents that happen to be working in parallel. These agents are a step above that. They coordinate with one another.
They talk with one another. They have a team lead. And essentially, they act as a real dev shop.
And that allows you to build more complex projects with the confidence that they're actually going to work. So, in this video, I'm going to show you how agent teams work, how to set it up yourself, and I'm going to run you through multiple real world examples of agent teams in action. So, with that, let's jump into it.
So, we're actually going to work backwards in this video, and we're going to start with the output that Cloud Code was able to give us using Asian Teams. That is what we're looking at right here. It was able to oneshot this dashboard, this internal dashboard for an AI agency.
But this dashboard is more than just, you know, a bunch of bar graphs. It was actually six separate modules. It put together using six team members and they all did their individual part, tied it all together, and produced this coherent app.
So obviously, we have our dashboard that we're seeing here. We have an entire client page that includes status and retainers. I can click inside the client, get more information.
I have a projects page, which is a whole cananban board. I can click in there, right? I can do subtask.
I can see time entries. I can go to the time entries page. I can have team members add how long they've been working on specific stuff.
We have invoices tied to the time page and a settings module. That's actually quite a bit, right? And if we look at the prompt I gave it, right?
A lot going on here, right? So, but the cool thing here isn't necessarily that it oneshotted it, which is pretty awesome. It's the fact that it was able to spin up six sub aents in parallel that are able to talk to one another that are able to execute individual parts of this larger app hole and put them all together all on their own.
That's pretty sweet. And again, this is a fairly complex thing to oneshot, right? All these different modules, all these things that need to come together.
So, that's what we're working with. How do we actually get to here? And we'll come back to this and I'll show you what it looked like at the end when I didn't use agent teams and what the difference is.
I think that'll be kind of surprising too. But I just wanted to show you this up front cuz I thought this was super cool. I was able to do all these things at once.
But now let's actually explain what the heck agent teams are and how they differ from this normal sub agent system we've been using now for several months. So we saw the final output. What's going on under the hood here?
Well, first of all, let's talk about what happens normally. So normally we have you and you have cloud code, right? And this is normally where most people just sit just in this little relationship.
But sometimes we want to use sub aents, right? Let's imagine I'm creating a project and I want that project to have a sub agent that works on the user interface that works on the back end and then I have one specifically for let's say databases, right? So UI backend databases, right?
When I spin up those sub aents in a normal context, what happens? Well, they go off and they kind of do their own thing. They don't talk to one another, right?
The UI sub agent is never going to talk with the database agent. Now, they'll go and talk with this main guy up here and feed information back, but essentially think of them as like freelancers, as mercenaries, right? You hire them on to do one specific tasks.
They go out and they do that task and then they come back and say, "Hey, it worked. It didn't. " Right?
Very much siloed. Now, let's compare that with agent teams and let's have kind of the same dynamic, right? I need one.
I need to do the UI. I need to do the back end and I need to also do databases. Well, what's different here?
Well, what's different? When I spin up an agent team, it's not only going to have these like sub aents that are doing those things, it's also going to create a team lead, right? We now essentially have a middle manager in the loop.
And so these sub aents, right, these team members will talk to and report to this team lead who coordinates everything and make sure everything makes sense in the greater hole. But it goes a little bit deeper than that because these sub aents can actually talk to one another. So database can talk to backend, backend can talk to UI, UI can talk to database, so on and so forth.
And so you actually have sort of a real team dynamic here where you have all these individual sessions of cloud code that have been spun up, but they can talk to one another. And what that means is we can actually work on much more complex projects and expect better outcomes versus the typical sub aent system which is much more attuned for like one-off tasks, right? Things that can be done in a vacuum, right?
I just need to send this guy off over here to the ether to do this one task and do it well and come back to me. So that big picture is how these things compare and the differences, right? Big difference.
teams can talk to one another. We have a coordinating team leader. These we're just off on our own.
We're doing our own thing, right? Also trade-offs. We need to think about this system, right?
This is better again for smaller tasks and it's also going to take less tokens because this agent teams dynamic obviously with all this interplay that just eats up tokens. Now, luckily for us, Anthropic gives us some guidance on where team sort of excels, right? And there's four big places.
This again comes direct from their documentation. and I'll put a link to it down below. So, teams excels in research and review, new modules or features, debugging with competing hypotheses, and cross-layer coordination.
Now, Enthropic also gave us this golden nugget where it compares agent teams versus sub aents in a little more depth, right? In terms of context, communication, coordination, what they're best for, and token costs. And this sort of just summarizes everything I've been saying.
Now, let's talk about actually enabling agent teams because they are disabled by default. This is an experimental feature that they will probably continue to build upon. You need to go into settings.
json and change this environment variable to one. So, here it is. If that's too confusing, which it probably is, or you're just lazy like me, all you need to do is go to the link to the documentation, hit copy page, drop it into cloud code, and say, "Hey, I want you to enable agent teams.
" It will do it on its own. And then you just need to restart the cloud code session, and you're ready to go. Now, one last thing before we get out of the documentation.
Let's talk about starting your agent team. So you went inside of cloud code, you told it, hey, change the variable, it's ready to go. To actually have it use agent teams, you have to explicitly tell it.
So they give an example here and it says create an agent team to do blah blah blah. One teammate here, one teammate there. You need to use that verbiage or something very close to it of create an agent team because it won't do it automatically, right?
You really need to prompt it. That's its trigger. So now we're inside of Cloud Code and we're going to do a quick demo of how to spin up the agent teams and what it's going to look like for you.
And then we'll do the side to side comparisons of the projects I did where I use agent teams and when I just use the normal cloud code setup with your standard sub aents. So here's the prompt. I'm designing a CLI tool that helps devs track to-do comments across their codebase.
Create an agent team to explore this from different angles. One teammate on UX, one on tech architecture, one playing devil's advocate. So I said create an agent team, right?
I use that specific verbiage and then I spelled out what each teammate should do. Now, if you don't know what each teammate should do, that's okay. Just put this in plan mode and say, "Hey, help me out.
How should we divide up this task for an agent team? " And now Cloud Co's telling us, "Hey, we have all three agents running in parallel. " Tells their name, UX, researcher, architect, advocate.
Gives us their role and their exact focus. And you can see here when it goes through the tasks of what it's doing, you can see which specific teammate was assigned that. So we can see analyze UX is to the UX researcher, right?
We can see down here also you can see all of your different um teammates. And if I do shift up arrow, I can actually see what they're doing. You can even send messages to the specific teammate.
So if I do at@ ux researcher right I now am giving a prompt to that specific teammate right so I could say hey make sure there is a settings page or something like that right so you actually have the ability to interact with these teammates it's not a complete black box and this is the terminal when I was doing my dashboard project the one you saw at the beginning you can see the same sort of pattern right specific task given to specific teammates means I have my all my teammates listed down here below. And at any time again, I just do shift. I just do the up arrow.
I can see what they're doing and I can actually send them messages. That's how the teams essentially UI should look to you as the user. It's pretty transparent.
And also note for this large project that it did one shot right at this point in time it was 23 minutes in. I think overall took about 30 minutes. And you can see we're sitting at 330,000 tokens.
So you know when you use teams you can have large token costs. So, I think this is definitely someone who is on some version of the Max plan. So, now that you understand what agent teams are and how to get it spun up inside of the terminal, let's actually do some sidebyside comparisons of the actual outputs.
Is Teams really putting us a step above what we would normally get. So, for the first project, I had Claude Code create an AI powered proposal generator. The idea was you talk with a client, you get some information, you throw it all into here and it generates a pretty nifty proposal, something they could actually present to a client.
And to make it easier, I will actually just turn off my camera from now. You can just hear me as a disembodied voice. So on the left, we have agent teams.
On the right, we have just your standard Claude. Left teams, right is standard. Right off the bat, pretty similar.
These were both oneshots. I didn't do any like crazy thing. They essentially had the exact same prompts minus the teams one saying, "Hey, do teams.
Here's what the teammate should look like. " So overall, first impressions pretty similar. A little bit more polish on the teams, right?
This UI does look a little bit better. Now, as we click through and we look at the generate proposal section, again, just a little bit more polish with this one on the left. And this wasn't a particularly difficult app.
This wasn't a huge ask. But the biggest differences is less about the functionality and I think more about the UI. And when we look at the actual summary that's produced, again, very similar, right?
I can't really say there's a huge difference between the two. If anything, the one on the left, the ROI projections for the team one looks way worse than the one on the right. So, the outcomes in this instance, a pretty simple ask, were virtually the same, right?
So, I decided to step it up a notch, which is where I did the dashboard prompt, the one we saw earlier, the one that had six modules and actually had quite a bit going on. And here's the two side by side again. So on the left, the teams on the right, no teams, just your standard thing.
And again, what's the big difference here? I think UI polish, to be honest, spoiler alert, I think UI polish is the biggest thing here. Although you do start to see the Teams version, again, the one on the left, pull away a bit, I think, as you add more complexity, you know.
Um, so we can go to projects. Camber and board looks good for both. Although I do really like how this one looks on the left.
We have our time sheets. Again, this just looks a lot better. On the left and even though we can kind of just handwave the UI stuff away to a point for anyone who's worked on these apps, the UI isn't a trivial task at all.
So, I will say the teams I think did a lot better on this internal dashboard in comparison to the normal version vice the first example. But again, is it necessarily a huge mind-blowing difference? Not necessarily.
And to be honest, I think that speaks more to the power of Opus 4. 6 than anything. What I would really like to do would be to give it even something greater to test.
um like a truly complex SAS project project, right? Um and I think that's probably going to be essentially part two of Asian teams in action that I would like to explore, but I still was impressed with what it was able to do. I mean, Opus 4.
6 as well. I think if I had Opus 4. 5 try to oneshot this, I don't think it would have looked nearly as good and still did a really solid job across the board.
And in terms of the token difference, yeah, teams was more. I get a little more polish. Um, but it wasn't like a massive token difference that you would expect, but again, it varies, right?
Because one agent team could have 20 teammates, one could have two. So, it also helps to know what we're actually dealing with. So, that's agent teams in a nutshell, guys.
I think it's super cool, especially when you start comparing these new features with a lot of the cloud code frameworks we've seen coming out lately, things like GSD, right? you start to see a lot of the principles from those, right? Sub agents, fresh context windows, and now we're doing sort of sub aent to sub aent communication.
Really great stuff. And I think kind of being on the bleeding edge of cloud code and instances like this is really important if you're trying to build complex applications. So, as always, let me know in the comments what you thought.
Make sure to check out all my resources in my school communities down there pinned in the comments. The free one, tons of free stuff. The paid one is more if you're someone who's serious about this AI stuff and is trying to figure out how can I actually turn this into a business and make money with it.
So again, let me know what you thought.
