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.

Greetings from Las Vegas — just landed for Oracle AI World 2025.

This week:
→ OpenAI drops three major announcements at DevDay 2025
→ IBM × Anthropic team up to accelerate enterprise Agentic AI
→ Andrew Ng (a.k.a. Mr. AI) launches a new course on Agentic AI

Plus: a deep dive into what Bloomberg calls the $1 trillion AI bubble — and why it’s not what you think.

Let’s dive in.

Weekly Field Notes

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

🌀 OpenAI debuts AgentKit
→ Visual toolkit for building and deploying AI agents. Connect MCPs, ChatKit widgets, and APIs.

🌀 OpenAI rolls out ChatGPT Apps
→ Third-party apps can now run inside ChatGPT — turning it into a full App Store for AI. Think plugins, tools, and workflows that live directly in your chat window.

🌀 OpenAI graduates Codex to GA
→ Adds SDK access and Slack integration. Codex quietly evolves into OpenAI’s enterprise code agent.

🌀 IBM + Anthropic partner up
→ Claude SDK integrated into IBM’s enterprise dev workflows — governance meets autonomy.

🌀 Anthropic releases Claude Agent SDK
→ Grants Claude computer-level autonomy — file ops, browsing, and full workflow execution.

🌀 Google unveils Gemini 2.5 Computer Use
→ Let’s agents click, type, and scroll on real UIs — a step closer to full autonomy.

🌀 Google DeepMind introduces CodeMender
→ AI agent that scans and fixes code security issues automatically.

🌀 ElevenLabs launches Visual Voice Agent Builder
→ No-code drag-and-drop builder for multimodal, voice-driven agents.

🌀 Amazon debuts Quick Suite
→ Internal agentic platform for BI and research — turns AWS data into autonomous insight flows.

🌀 Adobe introduces B2B Marketing Agents
→ Automates lead journeys, outreach, and pipeline acceleration.

🌀 Zendesk claims its AI agents now resolve 80% of support tickets
→ Customer support may become the first fully agent-run domain.

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

📘 Andrew Ng launches Agentic AI (on DeepLearning.ai)
→ Learn the four agentic design patterns — Reflection, Tool Use, Planning, and Multi-agent Collaboration. Includes hands-on labs and eval frameworks.

📘 Google launches Google Skills
→ A new human-centered learning platform built for the AI era. Integrates Gemini Code Assist (20+ hands-on labs), career-linked certificates, DeepMind-authored content, and new badges like vibe coding (seems to become a thing) and AI agent development.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 IBM + Anthropic publish guide on architecting secure enterprise AI agents
→ Full lifecycle blueprint covering trust, governance, model control, and safety architecture for agentic systems.

🔹 Stanford University proposes Agentic Context Engineering (ACE)
→ Self-improving LLM framework that rewrites its own context for better reasoning.

🔹 Samsung Electronics unveils Tiny Recursive Model (TRM)
→ 7 M-parameter model beating much larger LLMs on reasoning tasks. Efficiency > scale.

🔹 Air Street Capital drops State of AI Report 2025
→ Published annually since 2018, the report tracks key research trends, funding flows, and policy shifts shaping the next phase of AI.

🔹 Jacob Yates on AI for Science
→ Refreshing read with a strong message: The future of science isn’t AI replacing humans — it’s AI expanding how humans think.

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

Bubbles are hard to spot — and harder to time.
But they’re easy to explain.

A new technology raises expectations → prices rise → speculators pile in → everyone bets they can sell to someone else for more later.

This week, Bloomberg dropped a chart that nails what’s happening in AI right now — a tight loop of money, chips, and promises:

• Nvidia sells GPUs to AI labs → labs raise or pre-sell future value → they double down on GPU orders.

• Cloud providers sign mega-deals → labs give equity or options back to infra vendors.

• Startups orbit the big nodes (OpenAI, xAI, Microsoft, AMD) → pulling demand forward on the belief that “AGI margins” will pay the bill later.

The Loops Behind the AI Boom

Speed, scale, and signaling make the AI economy currently look unstoppable. But behind that momentum lies real fragility — every rapid move increases the risk of imbalance.

• Capex: Chips → Models → Hype → Capital → More chips.

• Revenue: Cloud credits → Usage → Growth → Bigger commitments.

• Power: PPAs → Capacity → Priority access → Moats.

• Talent/Tools: CUDA, Triton, vLLM, MCP → Lock-in → Switching costs.

If any loop breaks — capex, revenue, power, or talents/tools — the wheel wobbles. And when it does, the survivors will be those building for 2030, not 2025.

At first glance, it looks like an AI bubble — prices, hype, and capacity feeding each other. But here’s the key: the technology isn’t the bubble — the speculation is.

The infrastructure is real. The compounding effects are already in motion. A correction will come — and that’s healthy/needed. It won’t kill AI — it’ll kill the grift.

Beneath the hype, something much deeper is happening. The infrastructure of the 21st century is being built — silicon, power, compute, and data centers. New models, open weights, agentic platforms. And a new generation learning to build with AI, not against it.

That’s the real loop: capital becomes capability, and capability rewires civilization.

Yes — companies chasing short-term funding rounds will fade. But the ones building foundations — real infrastructure and lasting capability — will define the next decade.

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

I’ve been experimenting a lot with AI-coding assistants lately — and it’s great to see Atlassian stepping into this space with a serious enterprise spin.

Rovo Dev is Atlassian’s new agentic AI for developers — built around agentic workflows, a secure CLI, and full SDLC coverage.

What it does:
→ Standard + one custom experience for planning, coding, review (Bitbucket + GitHub), deployments, and CI/CD troubleshooting
→ Remote MCP Server for secure IDE or web access — without exposing source code
→ Granular permissions and enterprise-level governance
→ Deep integration with Jira, Confluence, and Bitbucket

How it works:
Rovo Dev runs via CLI or Remote MCP, giving you one task-specific agents that analyze code, generate commits, summarize deployments, and debug pipelines. Each agent operates contextually — and all stay within Atlassian’s trusted stack.

Try it:
→ Rovo Dev — general access just launched.