multiplayer AI agents are the future

Hello young padawans,

I’m in Copenhagen right now, and it’s easily in the top 5 cities I’ve ever been to. Super clean, feels safe, and everyone has immaculate style.

If you’re in the area, reply to this email - I’m thinking of hosting a little meetup.

Anthropic shipped Claude Opus 5 on 24 July.

It's half the cost of Fable 5 (same price as Opus 4.8)

That price doesn't survive real work though. From numerous user tests, it seems to run for longer and burns more tokens, so on certain tasks it often ends up costing you more.

Nate Herk ran both models on identical jobs in Claude Code. A landing page came in at $35.83 on Opus and took about an hour. Fable did the same job for $20.50 in 22 minutes, and he called the two results a tie.

Which is pretty much why I don't nerd out on benchmarks. With these models it's all about the vibe, and how they feel when you actually use them.

Every new model now is a bit like a new iPhone. The jumps from the 3 to the 4 to the 5 to the 6 were massive. The difference between a 14, a 15 and a 16 is minuscule.

Where Opus did win: bug-hunting in a real codebase. It scored it 93 out of 95 against Fable's 66, and it was cheaper doing it.

And on an audience-research task it worked through about 2,200 YouTube comments and 480 community posts, more thorough than Fable and cheaper again.

Where Fable still wins: anything needing taste or visual judgment. Carousels, slide decks, the look of a landing page.

The part I did find interesting is the “thinking levels” (this is a little slider you can play around with in most AI agent harnesses like Claude Code).

The lower ones seem to work better?!

Dan Shipper's team at Every spent a week testing and got better results on low and medium than on high.

And Anthropic's own numbers back it up: their chart has Opus 5 peaking at roughly 44.5% at xhigh thinking, then dropping to roughly 43.3% at max thinking.

Anthropic's advice is to start at '“high” thinking and step down. So if it's being annoying, try turning the thinking down before you go reaching for a different model.

We've all been moving from chat to agents this year.

But every agent people use right now is single-player. It's your agent, on your machine, doing your work.

It reminds me of using Microsoft Word before we could collaborate on files. You'd finish the thing, export it, email it to your team, and they'd send one back. Then Google Docs turned up and everyone just worked in the same place.

That's roughly where agents are now, and BUZZ is the first proper attempt at the Google Docs version. Block built it, Jack Dorsey's company, and it's free and open source.

It's a chat app that looks a lot like Slack, except your AI agents are actual members of it rather than bots bolted on.

You @ an agent like you'd @ a person, and it replies in the thread with progress updates as it works.

Each agent has its own identity, so you can see exactly which agent did what, the same way you'd see which colleague did what.

BUZZ is just the interface, and you pick what agent you plug into the back of it. Right now you can choose between Claude Code, Codex or Goose, with more expected.

So you're not locked into one company's AI, you BYO agent.

This is pretty neat: when you create a new agent and drop it into a channel halfway through a project and it reads everything that already happened. The history, the decisions, and the files. Then it carries on from there.

One agent can hand work to another, too, delegating to the rest without you managing it.

They also released a BUZZ mobile app on iPhone and Android, to talk to your agents on the go!

This app looks awesome, and I think this is definitely the future.

I tested it, and it's still in very, very early stages, so I'm not moving any of my real work into it for now. But I'll be keeping an eye on it, and I can see myself using it.

The main thing I didn't like: I couldn't really see and manage my skills in the interface, or my tools and context files.

Block is shipping fixes daily, so expect to see rapid improvements.

You record your screen while you do a task, talk through what you're doing as you go, and Claude turns that into a skill it can run again.

This feature lives in Cowork. It's in the + menu in the Claude desktop app, under Record a skill, on Pro, Max and Team plans.

The clever part is that it doesn't copy your clicks.

Instead, for each step it works out the outcome you produced. A file changed, a message sent, data fetched, etc. Then it picks the fastest, most reliable way to reach that outcome, which is often quicker than the way you did it.

So it’s not just blindly copying the mouse and keyboard actions you performed.

Talking while you work is optional, but it's the whole point. Narration is how it learns why you did something rather than just what you clicked.

Claude proposes the skill afterwards and you press Save or Dismiss.

OpenAI shipped the same idea in Codex on 18 June, about a month earlier.

A few things worth knowing before you try it:

Two things from me…

What they actually shipped on 21 July was three models. Two you can use, and one you can't.

Gemini 3.6 Flash is the everyday one. It scores the same as 3.5 Flash on the intelligence index, but is slightly cheaper and faster. So this is a speed-and-price release rather than a smarter model.

Gemini 3.5 Flash-Lite is the cheap, high-volume one, and the fastest model in the 3.5 family.

Honestly there are only two reasons I'd ever reach for a Flash model.

Speed, if you're building an AI product or app of your own and you want something quick with a big context window powering it.

And vision, if you're running a lot of visual analysis on images or videos.

That's about it.

Cloud agents are the future

I've been thinking lots this week about cloud agents.

right now, 90% of the agents we use are local agents, meaning they just live on our computer (like claude code). and it's become really apparent and clear to me that local agents are not the future.

Local agents have a few perks, such as it being able to take actions on your computer, but comes with significantly more downside…

There's no multiplayer available for you to use them with your team. If you shut your computer, your agent stops running, your scheduled tasks stop working, and you can't use them well from your phone either.

It makes so much more sense to have agents on the cloud, and I think this is now the next evolution you going to see unfolding.

Shopify store built entirely with Claude Code

I showed my buddy Jack how to use claude code, and he’s humming now.

He built his entire shopify store with claude. *chefs kiss

Check it out here

just wrapped up the 6th and final session of The AI Course last night, with 60 founders and business leaders.

most of them had never touched an agent before this.

Now they've onboarded their own agent team. written their own skills. and delegated their repeatable processes out to run autonomously.

Some of what they built:

👉 Sabe automated his ad publishing and got 2 hours a day back.

👉 Ben turned his whole client reporting process into one skill. half a day down to half an hour.

👉 Jai screen-recorded how he redlines contracts, and now an agent does it for him.

And that's 3/60.

Taking someone from "I don't even know where to start" to running their own AI team in a matter of weeks is the best feeling in the world.

A lot of you wanted in on this one and the timing didn't work, so we're building the self-paced version now.

more to come, stay tuned 👀