Hey, it’s Andreas.
Welcome back to Human in the Loop - your field guide to the latest in AI agents, emerging workflows, and how to stay ahead of what’s here today and what’s coming next.

This week we cover:
→ Major AI updates from AWS re:Invent 2025
→ IBM’s $11B acquisition of Confluent
→ Mistral open-sourcing the full Mistral 3 family
→ And a deep dive into the AI builder paradox.

Let’s dive in.

Weekly Field Notes

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

🌀 AWS re:Invent 2023 unveiled a wave of AI innovation.
→ During its flagship event in Las Vegas, AWS introduced the Graviton 5 processor alongside major updates across the AI stack, including Amazon Bedrock AgentCore, Trainium3 UltraServers, and new AI Factory infrastructure.

🌀 IBM to acquire Confluent for $11B
→ IBM is buying Confluent to build a smart, end-to-end data platform for enterprise GenAI and agents.

🌀 DeepSeek releases V3.2 and V3.2-Speciale
→ Two new open-source models that rival GPT-5 and Gemini-3-Pro on reasoning and coding benchmarks, using sparse attention to cut compute by 70% while delivering gold-level math and programming performance.

🌀 Runway releases Gen-4.5
→ Now the top-rated AI video model with a 1,247 ELO score, surpassing Google’s Veo 3, Kling 2.5, and OpenAI’s Sora 2 Pro.

🌀 Google Workspace adds workflow agents
→ Native agents can now draft emails, clean documents, summarize activity, and automate repetitive work inside Workspace.

🌀 Google rolls out Gemini 3 Deep Think
→ Deep Think mode for Google AI Ultra subscribers in the Gemini app available, with improved reasoning capabilities.

🌀 Microsoft launches VibeVe-Realtime-0.5B
→ Ultra-low-latency speech generation designed for real-time conversational agents and on-device interaction.

🌀 ByteDance releases Vidi2
→ A new multimodal video model that decisively outperforms GPT-5 and Gemini 3 Pro on major video-understanding benchmarks, delivering state-of-the-art retrieval, spatio-temporal grounding, and video QA in one system.

🌀 Mistral open-sources Mistral 3, including a new 675B Large model
→ Their biggest release yet. Full transparency across the family signals a strong push for open, high-capability systems.

🌀 Visa says nearly half of U.S. shoppers rely on AI for buying decisions
→ Agentic commerce is moving mainstream. Purchase paths are shifting toward AI-led comparisons and recommendations.

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

📘 DeepLearning.AI releases a new course on coding agents with E2B
→ Building Coding Agents with Tool Execution, created in collaboration with E2B.

📘 Overview: Top 12 MCP Servers to Know
→ A fast snapshot of the 12 most useful MCP servers spanning code, data, search, automation, payments, and collaboration.

📘 Overview: The Core Workflow of Vector Search and RAG
→ Great graphic that illustrates the full vector pipeline - from raw text to semantic embeddings, vector indexing, and similarity search - the mechanism powering RAG, agent memory, and modern AI retrieval.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 OpenRouter + a16z publish insights on open-source models
→ Joint report highlights where open models excel, how agentic inference changes infra requirements, and what “openness” means in the agent era.

🔹 McKinsey on the $2.9T AI Agent Economy
→ McKinsey’s new 50+ page analysis says AI agents could unlock $2.9T by 2030, as work shifts toward integrated human-agent-robot collaboration and AI fluency becomes the fastest-rising skill requirement.

🔹 Anthropic shows how Claude automates 60% of internal engineering work
→ Real-world workflow automation: PR reviews, migrations, refactoring, and internal tooling now handled largely by Claude-native agents.

🔹 Anthropic unveils AI Interviewer for user research
→ Scalable qualitative research through structured interviews and automated synthesis. Useful for product teams and model evaluation.

🔹 BCG on Building Real Enterprise AI Agents
→ BCG’s new 50+ page report breaks down why enterprise agents are hard to build, how to design and engineer them, and why the real bottleneck isn’t the LLM but the surrounding stack - data, IAM, governance, and legacy platforms that determine whether agents can scale reliably.

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

Everyone keeps saying they feel "behind" on AI. That feeling makes sense in a world where a new model drops every 72 hours. But the fear is misplaced. We are still in the early innings - and the real gap isn’t technical. It’s behavioral.

There’s a misconception that the people who build AI are the experts in how to use AI. They aren’t. As Ethan Mollick put it: "The people who build AI are often not the best at using it."

The bottleneck has moved from model builders to operators - people who can apply AI competently inside real work. But mastery requires a shift in behavior, not a new software certification. You can work at Peloton without being athletic; you can build AI without being fluent in its application. That fluency - AI Competence - has become the most in-demand skill in the economy.

Not another certification. Not another benchmark.
But the ability to:

• Turn messy, human workflows into agentic systems.

• Translate domain knowledge into prompts, constraints, and policies.

• Operate AI like a skilled engineer operates power tools: intentional, safe, efficient.

This is the hidden opportunity most people overlook. Every profession now has room for AI-athletes - practitioners who understand their field deeply and use AI to multiply their impact. AI doesn’t replace domain expertise, but it can amplify it and make you much more powerful.

NOW WHAT?

Stop asking where the ROI of AI is.

The advantage isn’t having AI.
The advantage is knowing where AI is useful and where it’s noise.

The value function is simple:

• Your value + AI

• Your expertise + AI

• Your unique perspective + AI

Use AI as a cognitive partner: to question assumptions, surface blind spots, and sharpen your judgment. If you’re still waiting for your tech team to “figure out AI,” you’re looking in the wrong direction.

Domain experts will define the next decade of AI progress - not the model builders.

Now is the moment to strengthen your core skills and learn how to pair them with AI. The gap hasn’t closed yet. But it will.

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

Microsoft dropped a free, open-source tool that lets you run AI models entirely on your machine.

Key traits:

• Zero cloud. Zero auth. Zero data leaving your device.

• OpenAI-compatible API built in

• Works on Windows and macOS

• Built for private, local-first agents

• Everything stays 100% private

How to install:

• Windows: winget install Microsoft.FoundryLocal

• macOS: brew install microsoft/foundrylocal/foundrylocal

You get a local inference stack that plugs directly into apps and agent workflows - without ever touching external servers.

You can try it here: foundrylocal.ai