Hey, it’s Andreas.
Welcome back to Human in the Loop — your field guide to what just dropped in AI agents, and what’s coming next in automation and real-world execution.

We just crossed 17,000 readers. Huge welcome to all the new faces. You're in good company.

Here’s what’s on deck this week:

• Microsoft turns MS Edge into an agentic browser with Copilot Mode,

  Anthropic drops 17 new Claude training videos (8h covering real Claude use cases), and LangChain launches Deep Agents.

But the big story?

• China just dropped its official AI action plan. While much of the West remains focused on frontier model performance, China is taking a broader, longer-term approach — building a fully integrated AI ecosystem spanning chips, patents, infrastructure, and STEM talent. In this edition, we’ll break down where China is gaining ground — and what it means for the global AI landscape.

Let’s get into it.

Weekly Field Notes

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

🌀 Microsoft turns Edge into an agentic browser
→ A new “Copilot Mode” gives Edge real-time memory, system commands, and goal execution — effectively turning your browser into a task agent. I’ve said it many times before: the browser is becoming the new operating system for agents.

🌀 NotebookLM adds AI-powered Video Overviews with visual synthesis
→ It now pulls in charts, diagrams, and even generates new visuals to produce animated explainers from your docs.

🌀 Meta on “Personal Superintelligence” + Billion-dollar recruiting push
→ In his new memo, Zuckerberg outlines Meta’s long-term vision. Meanwhile, Wired reports Meta tried to poach Thinking Machines Lab talent with $1B in offers — and was turned down. Signals both ambition and resistance in the race for top AI minds.

🌀 OpenAI adds ‘Study Mode’ to ChatGPT with Socratic tutoring
→ A new way to learn interactively. Ask questions, get challenged back. Great for deeper understanding and retention — not just passive answers.

🌀 LangChain introduces Deep Agents
→ These agents can plan, spawn subagents, use memory, and break down complex tasks autonomously. A strong shift toward recursive and self-directed systems.

🌀 Google updates A2A Protocol + ships Agent Toolkit
→ Pushes forward the standardization of agent communication and tool integration. Includes planning, routing, and execution modules.

🌀 Google drops LangExtract & RadExtract
→ New APIs for parsing structured info (tables, entities, timelines) from messy input — documents, chats, PDFs. Perfect tools for agents doing extraction and reasoning.

🌀 Firebase Studio adds agent templates for Flutter & React
→ Developers can now embed ready-made agents directly into apps. A big unlock for mobile and web-native agent interfaces.

🌀 Adobe brings reasoning agents into Acrobat + Experience Cloud
→ AI agents can now navigate complex document workflows — summarizing, cross-referencing, and reasoning across data layers.

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

📘 Anthropic drops 17 new Claude videos — 8 hours of GenAI learning
→ Includes tactical walkthroughs, best practices, and real-world innovation stories. Bookmark this one.

📘 Microsoft releases new course “MCP for Beginners”
→ One of the most beginner-friendly intros to the Model Contect Protocol. Covers everything from building your first MCP server to real-world deployment patterns. Includes also a 75-page cookbook and 10+ hands-on labs.

📘 Pipcat - Build real-time Voice AI agents
→ Good demo on how to create human-like, task-performing voice agents using Pipecat — an open-source, vendor-neutral framework backed by teams at NVIDIA, DeepMind, and OpenAI.

📘 DeepLearning course: Build Rich-Context AI Apps with MCP
→ This course walks you through building and deploying MCP-compatible apps.

📘 Google Gemini CLI Cheatsheet
→ Everything you need to turn Gemini into your personal SWE intern. Covers install, settings, MCP servers, slash/context/shell commands, and custom test-writing extensions. Highly bookmarkable!

📘 IBM crisp explainer on MCP
→ A beginner-friendly intro to Model Context Protocol — what it is, why it matters, and how it changes the way we build agents.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 OpenAI’s 120+ pages cookbook on real-world agentic tasks
→ Shows how agents handle evolving info over time — including memory, iteration, and external tool use. Tactical and very applicable for dynamic knowledge environments.

🔹 Microsoft shares 1,000+ GenAI use cases — with agents live in prod
→ One of the most detailed corporate catalogs to date. Covers agent deployments across HR, support, finance, and supply chain. Great source to inspire your own roadmap.

🔹 AWS on Agentic AI Frameworks, Protocols, and Tools
→ In a massive 70+ page guide, AWS breaks down the agentic AI landscape: from framework comparisons (LangChain, CrewAI, Bedrock Agents) to protocols like MCP and A2A.

🔹 Orb on AI Agent Pricing in 2025
→ In a report covering 66 vendors, Orb outlines the 8 core pricing models for AI agents.

🔹 Google’s list of 10 high-leverage agent use cases
→ From meeting prep to knowledge routing, this list shows where agents outperform humans on speed and scale. Good framing for pilot ideas.

🔹 Manus on Context Engineering for AI Agents
→ This deep-dive from Manus CTO shares battle-tested lessons: from optimizing KV-cache hit rates to file-based memory, masking vs. removing tools, recitation to focus attention, and the danger of over-patterned few-shotting.

🔹 OpenAI shares internal Codex playbook for engineering teams
→ Practical guide on how OpenAI engineers use Codex to debug, refactor, scaffold, and speed up delivery. Includes prompt patterns, setup tips, and real workflows.

🔹 Google Cloud whitepaper on Agentic AI for automation
→ Unpacks enterprise-grade agent architecture — from goal routing to tool orchestration. Highlights emerging design patterns and security implications.

🔹 Anthropic drops 'Persona Vectors' to track LLM behavior drift
→ You can now monitor how your model's tone, goals, and response patterns evolve over time. Important step toward explainability and safe delegation.

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

This week, China released its official AI action plan at the World Artificial Intelligence Conference 2025 in Shanghai — just days after the U.S. published its own.

The contrast couldn’t be clearer.

→ The U.S. frames AI as a race for dominance — focused on deregulation and growth.
→ China proposes global coordination, open-source development, and international infrastructure for joint R&D, AI literacy, and data sharing — especially for developing countries.

Chinese Premier Li Qiang even called for a new global AI cooperation body, warning that AI must not become an “exclusive game” for a few countries or companies. China is striking a fundamentally different tone — one that centers on inclusion, soft power, and long-term positioning.

And for many nations outside the Western alliance, this message hits differently.

But where does China really stand in the AI race?

Short answer: Catching up fast — and in some categories, pulling ahead.
But this isn’t a zero-sum game. It’s a multi-front race where different players lead in different lanes.

Here are five fronts where China is building serious momentum:

1. China’s Open-Weight Model Surge

While U.S. companies still lead on closed frontier models (GPT-4o, Claude 4, Gemini 2.5), China dominates the open-weight ecosystem.

Look at LMArena or Artificial Analysis — Chinese models now beat Meta’s Llama 4 and Google’s Gemma 3 in many benchmarks.

Chinese top performers:

• DeepSeek R1-0528 — strong reasoning and benchmark scores

• Qwen3-Coder — beats Claude Sonnet on multi-turn coding and tool use

• Kimi K2 — optimized for agentic workflows

• GLM 4.5 — released post-training software as open source

→ Insight: Open models are becoming China’s soft-power export — pairing capability with commercial freedom via Apache 2.0 licensing.

2. R&D Investment and Patent Power

China isn’t just building — it’s filing, publishing, and institutionalizing AI innovation at scale.

• ~28% of global R&D investment (vs. U.S. at ~29%), but with 3x faster growth

• 74.7% of global GenAI patents filed between 2014–2024

• Leading in NLP, computer vision, and robotics patents via top universities

→ Insight: Even if patent quality varies, the volume + velocity + vertical integration of China’s innovation machine is unmatched.

3. State-Coordinated Momentum

China aligns investment, infrastructure, regulation, education, and deployment — all as part of a national strategy.

• $139B+ in government-backed funding

• Strategic zones and public–private AI labs

→ Insight: Where the West debates regulation, China executes national policy. This is innovation at the speed of state.

4. Chips, Compute, and Deployment

Despite U.S. sanctions, China is building compute infrastructure at breakneck speed.

• Huawei CloudMatrix 384 — clusters hundreds of lower-power chips to rival Nvidia’s GB200

• Bloomberg uncovered a massive AI data city in the desert, reportedly set to run on 115,000 high-end Nvidia chips — despite bans

• Nvidia and AMD now cleared to resume chip exports after trade pressure and negotiation

→ Insight: Chip access is now a geopolitical lever — and corporate diplomacy is part of the game.

5. STEM Talent & University Divergence

China isn’t just producing more graduates — it’s building a long-term edge in scientific capability.

• 50,000+ STEM PhDs awarded in China in 2022 (vs. ~34,000 in the U.S.)

• China’s output has doubled since 2007; U.S. grew just 41%

• Undergraduate base: 39M+ students in China vs. 19M in the U.S.

But the divergence isn’t just numeric — it’s strategic:

• China is scaling up public funding for universities and R&D

• U.S. is cutting or freezing funding for top institutions (e.g. Harvard, Cornell)

→ Insight: Talent is a national asset. While the U.S. focuses on elite research quality, China is scaling the quantity + institutional backing to dominate core technology pipelines.

Final Thought

The U.S. still leads in compute, foundational research, and elite model performance. But China’s combination of scale, state coordination, open innovation, and deployment speed makes it a serious — and in some domains, dominant — contender.

AI isn’t one technology. It’s an entire stack — from chips to models to governance. And if you zoom out, it’s clear: China is not just building tools. It’s building an ecosystem

Yet, I don’t see this as a winner-takes-all race.

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

Sim Studio — drag-and-drop agent workflows, 100% open-source

If you’re building AI agents and want a clean local-first setup — Sim Studio is worth checking out. It’s like if LangChain had a visual editor and shipped with real-time execution + Ollama baked in.

→ Drag-and-drop UI (powered by ReactFlow)
→ Real-time agent execution and feedback
→ Native support for local models via Ollama
→ Connects to your stack — API calls, files, webhooks
→ Deploy however you like: NPM, Docker, or Dev Containers

How it works:

• Run npx simstudio for instant local launch (requires Docker)

• Or self-host via Docker Compose, Dev Containers, or full manual setup

• Works with your own models (GPU or CPU) using Ollama profiles

• Uses pgvector for embedding-based features like knowledge bases

Sim is a real platform for building and deploying agentic systems. Visual, local, and developer-friendly. Perfect for testing ideas, building internal tools, or shipping production-grade workflows with full control.

→ Try it now: Sim GitHub