echo-core/memory/kb/youtube/2026-02-07-claude-opus-46-agent-teams.md

Claude Opus 4.6: Agent Teams Change Everything!

URL: https://youtu.be/RWDK5414yL4
Durată: 11:54
Data salvare: 2026-02-07
Tags: @work @growth

TL;DR

Claude Opus 4.6 introduce "agent teams" - posibilitatea de a orchestra mai multe instanțe Claude Code complet separate (NU sub-agenți), care lucrează în paralel pe taskuri diferite. Demo: construire app global tide times cu 3 agenți (backend dev, frontend dev, skeptic/reviewer), apoi redesign cu 4 agenți competitivi (90s lover, Apple purist, Gradient King, Image lover).

Insights cheie

Agent Teams - Arhitectură nouă

• NU e vorba de sub-agenți într-o sesiune - sunt instanțe Claude Code complet separate
• Fiecare agent rulează în terminal propriu (tmux panes)
• Agenții pot fi specializați pe roluri diferite: backend, frontend, security review, etc.
• Colaborare prin fișier comun (claude.md) unde fiecare scrie decizii

Setup tehnic

• Feature experimental, trebuie activat în ~/.claude/settings.json
• Recomandare: tmux pentru mission control over agents
• iTerm2 pe macOS pentru split panes vizual
• Agenții se spawn-uiesc, colaborează, apoi se termină singuri când taskul e gata

Workflow demonstrat

1. Research - agenții caută APIs, frameworks
2. Plan - fiecare agent propune, skeptic reviewează
3. Approval - human-in-the-loop confirmă planul
4. Implement - agenții execută în paralel
5. Termination - agenții se închid singuri când e gata

Pattern interesant: Skeptic role

• Agent dedicat care NU scrie cod, doar reviewează
• Găsește blocking issues, strongly recommended fixes
• Joacă devil's advocate înainte de implementare

Multi-agent design battle

• 4 agenți cu preferințe estetice conflictuale
• Fiecare argumentează puternic pentru viziunea sa
• Team lead sintetizează și recomandă câștigător
• Human ia decizia finală

Upgrade-uri model

• 1 million token context window (poate citi codebase întreg)
• Better coding skills (standard la orice upgrade)
• Agent orchestration capability

Aplicabilitate pentru Echo

Ce folosim deja

• Sub-agenți prin sessions_spawn - similar dar diferit
• Izolare sesiuni pentru taskuri separate
• Colaborare prin memory/ și raportare

Ce am putea învăța

• Specialist roles: Backend, Frontend, Reviewer - în loc de agenți generaliști
• Skeptic pattern: Agent dedicat review/security care NU execută, doar critică
• Shared state: claude.md echivalent - poate memory/agent-shared.md?
• Self-termination: Agenții știu când au terminat și se închid singuri

Diferențe importante

• Agent teams = instanțe separate Claude Code (heavy)
• Sub-agenți Echo = sesiuni izolate în același gateway (lighter)
• Agent teams = best pentru coding projects mari
• Sub-agenți Echo = best pentru taskuri diverse (email, rapoarte, insights, etc.)

Transcript complet

[... transcriptul integral ...]

Claude Opus 4.6: Agent Teams Change Everything!

Claude Opus 4.6 just dropped and it changes everything. And I'm not just talking about prompting anymore. I'm actually going to orchestrate entire teams of separate agents to do a task for me in this video today. Today I'm going to go back to something I've been trying to build for a while. I'm going to show you how to build the ultimate global Tide Times app. And I'll even pit four agents against each other to battle it out to get the best design for my app. Now let's get into the good stuff. Claude Opus 4.6. And yes, this was announced from Anthropic just the other day. We're upgrading our smartest model. It's now Claude Opus 4.6 and improves on coding skills of course, every time a new model comes out, it's better. You just want to know what's good and what's happening and what can I do with it. And I can tell you that it's got better at two things. In my opinion, that I think are really, really cool if you're making apps and vibe coding, and that is a million token context window, that's incredible. It means it can read much more of your code base and work with them, But I think the most exciting thing is that you can now orchestrate teams of Claude Code sessions. That's right, multiple Claude Code instances. So we're not talking about one clawed code agent spinning up sub agents. We're talking about one Claude Code agent spinning up completely separate versions of Cluade Code. So you can just have tons of different Claude Codes running working on different things. It's called agent teams. And we're going to investigate it as we build this app. Okay, I've read all the docs so you don't have to. Now this is an experimental feature so you'll need to enable it. Now I'll show you how to do that in a moment. And also you can choose a display mode. Now this is really interesting because you can split it out into panes and make it look really sci fi and futuristic. and the documents actually tell you to use tmux to view your agents, because it's a well known command line tool that lets you run multiple sessions. And interestingly enough, it says to run it inside. Item two. Just the other week I did a video about my vibe coding stack, and I'm using terminus quite a lot to split out terminals, but I've never tried iTerm2, which is a terminal emulator for macOS, and it also says here on the description, it does amazing things and I'm kind of excited. So I'm going to install it and try and set up Claude agent teams inside. To get this working I'm going to edit the dot Claude directory and the settings dot Json file. And I simply go down to the bottom and I can copy and paste this code in to the settings file. This means every time I run cloud code it's going to work. But to start with, I'm not going to do that. I'm actually going to go ahead and brew install tmux on my Mac. This is going to enable me to have that mission control over my agent teams, which I really like. Okay. Now that's installed on tmux Dash CC, which is exactly what I'm told to do right here in the documentation. And you'll see I've got a tmux terminal. So inside my project folder I've made the tide times directory. Ultimate tide times here. And now we can run Claude. So I don't have to keep hitting buttons Claude dangerously skip permissions is what I'll run. And look at this. We are in. All right. Let's see what this can do. I want to build a global tide and sea temperature web app. Creates an agent team with three specific team mates to work in parallel. Now I'm going to tell Claude Code what those agents will do. Number one will be the back end dev focusing on integrating the global tide API and sea temps. And then number two. That's the front end dev who will build a responsive dark mode web UI that visualizes the data nicely. And then number three, and this is the important thing about agent teams inside cloud code. We're going to have the role of the skeptic, a security and UX researcher who plays devil's advocate. Then all right, follow these rules. And then number one will be require plan approval for backend dev and frontend dev before they write any code. That's important. Number two the skeptic should not write code, only review and plan docs. all teammates must update. Claude Dot, MD. This is really important with all their decisions. So I get visibility as my agent team's build. So that's my first prompt. Let's hit enter on Claude Code running on opus 4.6 and see exactly what happens. Time for a little coffee. Empty project directory. So it's creating the claude.md file. This is what I'd expect. Now let me create the task list. Now this is really good. You'll see I've still got the GSD plugin installed. I was using that to plan things out, but now it seems more and more with each update. Claude code is actually writing the plan itself. It has to research first, then design and plan implement. This is absolutely fantastic. Okay. This is really, really cool. Now you can see it's just popped open another terminal window here which is essentially popping open new agents and it's popping open another terminal here. This is truly futuristic, but I'm getting some security warnings as I'm running basically tmux inside item two. This is the first time I've experienced this. What I can see here is it says, wow, now there's three terminals here. What have we got? This is absolutely insane. Okay, so Claude will be able to read and edit. Yes I trust this folder. I'll say absolutely and the same on this one. So now we can actually see if we look over here, we've got a back end dev up here. We've got a security and UX review task going on over here. And we've got a frontend dev that I need to give permissions to as well. And they're all going to go ahead now you'll see my front end version of Cloud Code is actually running its own two sub agents as well. At the same time, working on best front end frameworks for 2026 charting libraries. So I'm going to get a really good looking app we can see up here. We've got more web searching going on here and more team approval being waited for. So I'll just say yes to all. And it's really cool to see that we've got app back end dev we've got at the skeptic, and we've got app frontend dev all doing their own separate tasks. over here in the main control panel we can actually see the to do list getting ticked off. And we can cleverly see the cloud code has actually said what agents it's waiting for to complete the work before it moves on to the next steps. The back end agent seems to have finished it says some research. It's designing the back end plan based on these findings. And of course it's going to pass that to the skeptic to see whether it gets approval or not. Now, while we wait for this to get done, I've just popped open another terminal window and I'm going to brew install glow on my Mac. And then we'll get into my projects directory and the ultimate Tide Times directory. And now this is great. I'm going to type in glow and then claw.md and look at this. Now we've got a nice looking file that is all marked in different colors. It's taken the markdown and made it look really nice. So I can actually view the full text that all my different agents have written in here. On how they're building the app. It's interesting to see how my skeptic has worked. It's read all the documents. It's found some blocking issues, strongly recommended and some nice to have fixes. And it's putting all of this into Claude, MD Okay. We can see my agents are now shutting down. It's just the back end dev that's left. And if we read to the bottom here, we can see the skeptic confirmed and terminated. So lots of my agents are basically agreeing. All is done. This is exciting. Once they're done, they wrap up and close their terminal windows. Health futuristic. Is that. And we get to the bottom now with the main terminal we can see the tide app is fully shut down and cleaned up. All three teammates terminated themselves. That's raw. The final isn't it? The project is ready to run both front end and back end to build, tested and documented. Jeanne-Claude, MD And it's popped open my ultimate Tide Times app. Let's type in La Hoya and see what comes up. There we go in California and see wow, okay. It's got the sea temperature of 16.6 Celsius. Cool. So I can flip over to Fahrenheit if I prefer. It's got this beautiful tide graph which is pretty accurate, and the tide times and what exactly to expect. It's done a pretty decent job, and it just worked out of the box with the first prompt, which is pretty impressive. Southend on Sea England United Kingdom and boom look at that. That's a pretty comprehensive graph even going into the minus numbers, which is really cool. What about trying Paphos in Cyprus? And yes, it appears there and indeed it seems to be absolutely correct. A very minimal tide change, a correct sea temperature here. We now have a global tide times up with sea temperature included, with just one prompt, three agents and a final prompt to get it working globally. Now let's have some fun with new agent teams. I'll say creates a team with four agents to improve the design of this app. Agent one loves 90s websites. Agent two loves Apple best practices. Agent three loves gradients on everything and agent four loves using images. They should all argue strongly for their case. This is going to be great. Now let's unleash the chaos spawning all four agents in parallel. We've got the retro 90s agent. We've got the Apple purist agent that's fired up. We've got the team lead over here that will ultimately have to decide things. We've got the gradient thing firing up here. This is absolutely hilarious. And one more firing up into action. The image lover. This is going to be a really colorful, beautiful and hopefully Apple best practices designed app by the end of this debate. All right. The 90s one says I've seen the full extent of this sterile corporate soulless. This is absolutely brilliant. Image lover has fired the first shot at gradient King debates that heating up Apple purists is going after Gradient King two and pool Gradient King is getting double teamed. Let's see how they respond. So do I go for the gradient king with the aurora backgrounds and animated ocean gradients. Zero bandwidth. Strongest debate performance apparently. So that is the one that would win. But I ultimately is the human in the loop. Have the final say. Then number two the image lover would have full bleed ocean photography, glass morphism columns, and cinematic immersion. If you want to follow the Apple purist, it's going to be SF Pro Frosted Glass, true black Hole retro 90s monkey Comic sans visitor counter wearing Ascii art pure joy and chaos. So let me know in the comments down below. Now to those four designs. Which one would you pick? well, seeing as it tells me the strongest debate performance came from Gradient King, I'm going to accept Gradient King this time and we'll have a look at what it came up with. And here is our new gradient heavy design. Yes we can see as I hover over things cards pop out. Everything has a gradient on. I can actually change this to somewhere like this in France. And we'll get the tide times. And everything is looking great here. It's fully functional, it's gradient heavy while not changing the design that much. It's definitely updated. I'm kind of wondering now what the 90s design would have looked like. And it's updating in real time. That's a proper 90s search bar. Let's type in Bristol in England. And yes, we get a very glowy graph now, full 90s style right there in front of me. So there you have it, Claude. Opus 4.6 is the best coding model yet. Yeah, potentially. Agent teams. Incredible. They can work for you overnight. Absolutely. Run it in item two with tmux to have multiple windows and keep visibility on every working agent. And be careful with security if you're running open claw. Lots to think about this week. I've really enjoyed doing this video. I hope you've enjoyed watching and got something out of it. Do leave a comment down below. Let me know what your favorite design would have been and join me in my free community that's also linked up down below too. Thank you so much for watching. And YouTube is showing a video on your screen now. You should watch next. Thanks.