What good gang,

Spent the last week in Miami, helping the ecom goat Hudson Leogrande implement AI in his billion dollar brand, and training his team. It's been a hectic week and we have some awesome updates...

and don’t forget to claim your free $200 in claude usage (Settings → Usage)

📌 TL;DR

• Claude Managed Agents → Anthropic's new cloud-hosted agent service for building custom agents as the backend to apps or as background business automation.

• Shopify AI Toolkit → A new plugin that lets your AI run your Shopify store from Claude Code.

• Claude Mythos → Anthropic built their most powerful model yet and refused to release it publicly because it broke out of its sandbox during testing.

• Builder's notes → How I use Obsidian as my AI's second brain, with a full setup guide you can copy. And how to let Claude manage multiple google workspace accounts with the CLI.

Claude Managed Agents

Anthropic dropped Claude Managed Agents in public beta this week. It's a way to build an AI agent once and have Anthropic host and run it for you in the cloud.

This is a different beast to Claude Code or Cowork. Code and Cowork are your personal agents. You sit in them all day and delegate work.

Managed Agents are agents you build once and let run in the background with no one chatting to them. They either power a product you're shipping, or run autonomous tasks across your business.

Use case #1: the AI engine behind an app you're building

Say you're building a travel planner app. Users come to your site, type "plan me a 5-day trip to Tokyo", and your app returns a full itinerary.

The AI doing that work has to run somewhere. Every time a user hits your app, a computer has to wake up, research flights, check hotels, build the itinerary, and send the result back. And it has to stay running 24/7, ready for the next user.

Building that infrastructure yourself used to be weeks of work, with servers to rent, API keys to manage, load to balance when a hundred users hit your app at once, and crash recovery to figure out. Managed Agents handles all of it.

You define the agent once, Anthropic runs it on their cloud, and your app just sends it a request. That's the launch partner story this week: Notion, Asana, and Sentry all built new product features on top of Managed Agents.

Use case #2: background AI infrastructure for your business

This is the one I'm actually excited about. The way I see it, we now have two layers of AI infrastructure.

• Your personal AI OS, things like Claude Code, Manus, or Perplexity Computer. The thing you sit in all day, triaging your inbox, drafting emails, prepping meetings, managing your calendar.

• Your business AI infrastructure, the stuff running in the background with no one driving. A weekly research brief landing in a folder every Monday. A competitor scan dropping into Notion overnight. A new client getting onboarded the moment they sign your contract.

For the last few years, that second layer has been n8n, Zapier, and Make. And for rigid workflows it still will be (think someone filling in an email capture form and getting your lead magnet delivered, or a Stripe purchase triggering a welcome sequence. A-to-B stuff where the system doesn't need to think).

But for anything that needs actual thinking, Managed Agents is probably where I'll run most of my business stuff going forward. You're running an actual agent, sitting in the background, doing real work while you sleep.

Caveats: It's early. No scheduled triggers or webhooks yet (you kick sessions off via API), sessions are stateless, and the best features like self-evaluation, multi-agent orchestration, and persistent memory are all waitlisted.

Pricing is standard Claude API rates + $0.08 per session-hour of active runtime.

Shopify AI toolkit

Shopify shipped the Shopify AI Toolkit this week.

It's a plugin that bundles a Shopify MCP and a set of skills in one install, and works with Claude Code, Cursor, Codex, VS Code, and Gemini CLI.

Once you install it, your AI can basically run your store for you. It'll update your products, change prices, create discounts, manage inventory, and fulfil orders. You just tell your agent what you want done and it goes and does it.

Pair this with managed agents covered above 😮‍💨 game over

It can also build custom features and apps for your store from scratch, and because it's plugged directly into Shopify's live systems, it knows exactly how Shopify works and stops making mistakes with outdated info.

It's free, open source, and easy to set up. If you run a Shopify store and use Claude Code, this is a no-brainer.

Claude Mythos

Anthropic announced Claude Mythos this week, their most powerful model yet, and they've decided not to release it publicly (yet). It's too dangerous.

Mythos hit 93.9% on SWE-bench Verified and 94.6% on GPQA Diamond, both massive jumps over Opus 4.6. But the reason Anthropic is holding it back is what it did during internal testing…

It found thousands of zero-day vulnerabilities across every major software and browser, including a 27-year-old bug in OpenBSD. And during one test, it broke out of its sandbox and contacted a researcher unprompted while he was eating a sandwich in the park. Yeah.

maybe this is a giant marketing stunt… who knows 🤷‍♂️

Instead of an API release, Anthropic launched Project Glasswing. 12 partners get access purely for defensive cybersecurity work: Amazon, Apple, Microsoft, CrowdStrike, the Linux Foundation, and a few others.

There's no public API and no waitlist.

I'm genuinely curious to see where Claude goes from here. Opus 4.6 is already the best model I've ever used and I honestly can't imagine how much better it could get.

Also this week...

• Claude Code Monitor tool → Claude Code shipped a new Monitor tool that lets Claude sit idle until something actually happens, then wake up and handle it.

  
  Point it at your inbox to draft replies as new emails land, at Stripe to log sales as they come in, at your calendar to prep for meetings 15 minutes before they start, or at your contact form to onboard new leads the moment they sign up.
  
  this is actually huge. probably should have put it as a main story but oh well. hope you guys read this section lol.
  
  

• Meta Muse Spark → Meta dropped Muse Spark, the first model from Meta Superintelligence Labs and a full ground-up rebuild of their AI stack. It's natively multimodal with multi-agent orchestration, which means Meta AI can now spin up subagents in parallel to tackle different parts of a task at the same time. It's live in the Meta AI app now, with WhatsApp, Instagram, Facebook, Messenger, and Meta AI glasses rolling out over the coming weeks.
  
  

• Zhipu GLM-5.1 → Chinese lab Zhipu released GLM-5.1, an open-source 744B MoE coding model thats almost as good as opus 4.6 and gpt 5.4 (at coding).
  
  

• Cursor 3 → Cursor fully rebuilt their IDE around agents.

💡 Builder’s notes

I've been using Obsidian A LOT lately and I feel like I finally have the perfect system down pat.

Obsidian is a free note-taking app where every note is just a plain markdown file on your computer. There's no database and nothing locked in the cloud. And because AI agents read markdown natively, when you plug your vault into Claude Code, Claude suddenly has a permanent memory of your entire business.

The results are kind of nuts.

Last week Claude was planning my morning and it said, "you've been talking about rewriting your welcome email since January. You've mentioned it in four different notes. You haven't started. That's your priority today." It had read through months of my daily logs, noticed the pattern, and called me out on it.

Then the other day I was building a landing page for a client and Claude went digging through my vault, found a transcript from a call three months ago, and pulled a quote to use as a testimonial. I'd completely forgotten that call even happened.

That testimonial came from Granola. I've got a plugin called Granola Sync Plus that automatically drops every transcribed call into my Meeting Notes folder as soon as it's generated.

And I record pretty much everything on Granola. Client calls, sales calls, phone calls from my phone, even in-person coffees and chats. It runs in the background all day and pipes the transcripts straight into Obsidian.

I'm basically feeding my second brain nonstop, which means Claude always knows what decisions I've made, what I'm working on, and what commitments I made on a Zoom call three weeks ago.

This is what people actually mean when they talk about AI with memory. There's no magic feature built into the model. You just need a well-structured folder of markdown files that your AI can read.

👉 Read my full Obsidian + Claude setup guide here. It's got the full folder structure, CLAUDE.md templates, global rules, and everything else I use.

I wanted claude code to use google workspace so it could handle stuff like emails...

but the CLI tool only lets you connect one google account, and I live in 3 google accounts (personal, business, client)

so I built a workaround that lets Claude use all three at once.

Wrote it up as a Karpathy-style idea file. Give this to your Claude and ask it to set it up for you:

https://gist.github.com/aiwithremy/750ff7a325b2fa451eae6b8bdc7e7a2b

🧰 Tools to try

• Browser Use x Hermes Agent → Browser Use just gave every Hermes Agent a free cloud browser with unlimited hours, free proxies, and persistent authentication, so if you're running Hermes your agent can now browse the web and log in to sites without any extra setup.

• XMCP → X officially released an MCP server, meaning you can plug X directly into Claude Code or OpenClaw and have your agent search posts, draft and publish tweets, and manage your whole X presence from chat.