Hey, it’s Andreas.
One of my big predictions for 2026 was that agentic AI and human agency would take a real leap forward. Looking at the latest releases around Claude Code, and now this week with Claude Cowork, this is playing out faster than expected. We are clearly moving from demos and early adopters toward mainstream agent usage.

Today, we take a closer look at Claude Cowork and what it signals for the next phase of AI adoption. And we also cover:
→ Meta’s new Meta Compute initiative
→ OpenAI moving toward ads in ChatGPT
→ Thinking Machines losing another key co-founder

Let’s dive in.

Weekly Field Notes

🧰 Industry Updates
New drops: Tools, frameworks & infra for AI agents

🌀 Meta and Zuckerberg launches “Meta Compute”→ A new top-level AI infrastructure initiative targeting tens of GW this decade.

🌀 OpenAI will bring ads to ChatGPT → Sam Altman says ads will be tested for Free + Go tiers, with conversations kept private from advertisers and higher tiers staying ad-free. 

🌀 OpenAI and Cerebras sign multi-billion dollar compute deal → With the goal to create ultra-fast inference for agents at scale.

🌀 Thinking Machines Murati loses co-founders back to OpenAI → Thinking Machines parted ways with co-founder Barret Zoph. It’s the third co-founder exit in under a year.

🌀 Anthropic launches Anthropic Labs led by Mike Krieger → A dedicated unit for experimental AI products.

🌀 Google launches Universal Commerce Protocol → Aims to standardize how AI agents browse, compare, and purchase products.

🌀 Google ships Gemini “Personal Intelligence” → Connects Gmail, Photos, and YouTube into a unified personal context layer.

🌀 Microsoft outlines a 5-point plan for community-first AI data centers → Focus on local engagement, sustainability, and shared infrastructure.

🌀 Zhipu AI releases GLM-Image → Open-source image generation that runs on consumer GPUs and beats Google in some benchmarks.

🌀 Replit introduces Vibe Coding for mobile apps → Natural language to native mobile builds. Extends vibe coding beyond web and into app ecosystems.

🎓 Learning & Upskilling
Sharpen your edge - top free courses this week

📘 9 FREE AI courses → A curated stack across GenAI, ML fundamentals, deep learning, foundation models, prompt engineering, and RAG agents. If you do one thing in 2026: block 1-2 hours weekly to learn, experiment, and apply AI hands-on.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹  DeepMind CEO Demis Hassabis on CNBC’s “The Tech Download” → As always, a strong guest - with candid views on the AGI race.

🔹 AI Futures Project updates AI 2027 model + ships an interactive site → New unified timelines/takeoff model is anchored in METR long-horizon coding trends, and it pushes “full coding automation” out by ~3-5 years into the early 2030s. 

🔹 Cursor publishes a practical guide to coding with agents → The meta-lesson: the harness is the product - plan first, give verifiable signals (tests/linters), and manage context intentionally so agents can iterate reliably across larger changes

🔹 Cursor uses hundreds of GPT-5.2 agents to build a browser → A real-world stress test of agentic development at scale.

🔹 BCG on scaling enterprise AI agents → Building one agent is easy. Scaling agents across a company is the hard part - and BCG packs the “how” into a tight 54-slide playbook.

🔹 LangChain shares best practices for multi-agent architectures → Clear guidance on coordination, memory, and failure modes.

♾️ Thought Loop
What I've been thinking, building, circling this week

The biggest news this week came from Anthropic: they shipped Claude Cowork. 

Most coverage framed it as “Claude Code for non-developers.” True, but the more important story is distribution.

Claude Code is already elite. It’s Anthropic’s terminal-based agent that can read and edit local files, run commands, and execute multi-step work - most people use it for coding, but it’s powerful for far more than that. The limitation here was never capability. It was the interface. If your agent lives in a terminal, you’ve implicitly capped adoption to people who are comfortable with CLIs, file paths, and local environments. That is not marketing, HR, finance, business or ops. That is NOT most of the enterprise.

Cowork removes the terminal as the gatekeeper. It brings the same agentic architecture into Claude Desktop, where you grant access to a specific folder and Claude can plan and execute multi-step work inside that boundary.

What Actually Is Cowork?

In a normal Claude conversation, you chat back and forth. Claude suggests things, you copy-paste, you do the work.

Cowork flips this model.

It’s the Claude Code agentic core inside Claude Desktop, exposed through a familiar chat UI instead of a terminal. You give it access to a specific folder, tell it the outcome, and it plans and executes: reading files, creating new ones, editing content, and organizing folders. The difference is agency - it doesn’t just recommend actions, it takes them, and checks in before major steps. You can assign it a task, go for a coffee, and when you return, the task will be completed.

In short, it’s Claude Code repackaged for general productivity, now accessible without a terminal. Here’s the flow:

1. You write a prompt

2. You point it at a folder

3. Claude reads, edits, or creates files in that folder to complete the task

Why the Shift From Terminal Matters

Claude Code has become a favorite among technical builders - and in my view, there is currently no comparable tool at this level. But it is gated by one thing: the terminal. That single interface choice creates a quiet distribution ceiling.

Most of the workforce will not install CLIs, navigate file paths, or debug local environments just to get value. For many, the terminal remains a psychological barrier:

• a black screen

• no visual UI

• no obvious undo

• file paths and CLI muscle memory required

So you end up with a weird gap: a tool that can do real work - organize files, process data, ship artifacts - but only a small slice of people ever touch it.

Cowork is Anthropic’s fix.

Same power, different interface. Claude Code was built for builders. Cowork is built for everyone else - which, in enterprise terms, is most of the company. Claude Cowork completes work like this with much more agency than you’d see in a regular conversation. Cowork is an early signal of what’s coming next: agents embedded directly into the surfaces where non-technical teams already work, with the ability to plan, act, and ship outputs without touching a terminal.

PS: If you are non-technical, I would still encourage you to try Claude Code directly. It takes time, but once you learn the basics, the terminal stops being intimidating and starts becoming leverage.

🔧 Tool Spotlight
A tool I'm testing and watching closely this week

I came across ActForge this week - and it’s already saved me a lot of time on the annoying “small-but-many” tasks: cleaning files, pulling info from apps, and turning messy inputs into structured outputs. It’s a smart desktop wrapper built on top of Claude Code and feels like an extended version of Claude Cowork.

What it does:

• A desktop-native autonomous worker, not a chat-only assistant

• Operates across local files plus apps like Gmail, Drive, Slack, Chrome

• One-click integrations - no API glue code

• Plan Mode that turns vague asks into executable steps

• Built on Claude Code foundations, but usable without living in a terminal

How it works
You connect your apps and give it a folder. Then you describe the outcome. It plans and executes - reading, creating, organizing, and navigating with real agency. 

Try it
Free download for Mac (requires Claude Code).