#3 Edition: Stone Age to AI Agents — This curve doesn’t flatten

Hey, it’s Andreas.

huge welcome to the 2,315 new subscribers who joined this week. Big moves, big momentum. Let’s get into it.

This week’s edition covers:

• OpenAI unlocks enterprise data flows inside ChatGPT — big move for integrated agent workflows.

• French startup H Company just came out of stealth — launching a modular platform built for agent collaboration (European answer to Manus AI).

• and much more…

Let’s dive into it!

Weekly Field Notes

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

🌀 OpenAI added Connectors to ChatGPT — real-time access to Google Drive, GitHub, and SharePoint inside the chat.
→ Finally makes ChatGPT actionable with enterprise data. Big step for integrated workflows and autonomous systems.

🌀 Mistral launched Mistral Code, their enterprise-grade AI coding assistant.
→ Goes beyond completions: autonomous debugging, doc writing, testing — all inside your IDE. Built on Codestral + Devstral.

🌀 Cursor 1.0 dropped with dev-first agents like BugBot (automated PR reviews) and a Background Agent for async workflows.
→ Cursor is now more than vibe coding — it’s becoming a serious IDE for agent-native development.

🌀 ElevenLabs released Conversational AI 2.0 — a real-time voice agent platform.
→ Handles RAG, outbound voice, and full-duplex dialogue. Strong move toward multimodal enterprise agents.

🌀 AgenticSeek — is gaining steam — a local, open-source alternative to Manus for browsing, coding, planning.
→ Lightweight, no cloud dependency. 16k+ stars already. Worth testing for local-first agent workflows.

🌀 Zapier now runs 800+ AI agents — more than their employee headcount.
→ Not mind-blowing agents and rather smart AI automation workflows, but still a milestone. They shared all internal templates here.

🌀 FDA launches Elsa — its the first AI agent worldwide for an entire agency.
Elsa helps staff review protocols, summarize data, and generate code — all inside a secure GovCloud setup.
→ If the FDA can scale agents, securely and fast — why can’t every other big company as well?

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

🟦 Google Gemini 2.5 - AI Engineering Workshop
Taught by Phil Schmid (Google DeepMind)
→ Practical intro to Gemini-based agents — covers multimodal input, function calling, and Model Context Protocol. Fully runnable on Colab.

🟦 Databricks - DSPy: Build & Optimize Agentic Apps
49-min course on building structured GenAI apps using DSPy.
→ Covers Predict, ReAct, ChainOfThought modules, MLflow debugging, and prompt optimization. Perfect if you're scaling LLM workflows.

P.S. Got a good new course? Send it my way on LinkedIn or just hit reply.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 Apple: New paper “The Illusion of Thinking” shows LRMs fail on high-complexity tasks.
→ When it gets hard, they stop thinking. Big implications for AI safety and reasoning hype.

🔹 Sakana AI introduced the Darwin Gödel Machine — a self-improving AI agent that rewrites its own code.
→ Outperforms hand-built agents on benchmarks. Not AGI — but real steps toward autonomous system evolution.

🔹 Meta open-sourced LlamaFirewall — a guardrail framework for agent security.
→ Blocks jailbreaks, misaligned goals, and insecure code. We need more of this.

🔹 BCG reports that 93% of execs say AI cuts costs — but only half see results.
→ Agentic AI delivers best ROI in content, customer support, supplier negotiations, and field ops. Key read if you're tying AI to real savings.

🔹 Cisco: New report on AI in customer experience.
→ Themes: proactive support, hybrid IT agents, ethical automation, and keeping empathy in the loop.

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

We humans aren’t built to intuitively understand exponential growth — especially in AI, where development isn’t just fast, it’s accelerating.

But zoom out far enough, and the pattern becomes clear.

Max Roser’s timeline helps put things in perspective:

→ Stone Age → Farming: 100,000 years
→ Farming → Steam Engine: 12,000 years
→ Steam Engine → AI Agents: 200 years

Read that again.

For 3.4 million years, we shaped stone. And 99% of human history happened before the steam engine.
And now? The internet, smartphones, and AI all arrived in the last 0.01% of that timeline.

From 2000 to 2014 alone, we experienced what would’ve counted as a century of progress. Moore’s Law predicted a 32× jump in compute — AI chips delivered 1,000×.

Children born today will likely see the 2200s — in a world completely unrecognizable to us.

This isn’t just exponential growth. It’s compounding.

And the implications are huge.

We’ve entered a phase where technology evolves faster than our institutions, schools, and governments can keep up.
AI agents, autonomous systems, self-improving tools — they’re no longer future ideas. They’re becoming part of daily life.

This shift will break old structures.
But it will also expand what’s possible.

New industries will rise.
Entire careers will be redefined.
Human potential will be augmented.

I believe the next 10 years will offer more opportunity than the last 100 combined.

Don’t fear the shift. Learn it. Build with it. Embrace it.

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

I am keeping a close eye on emerging European AI agent stacks — and this week, H Company (Paris, $220M raised) came out of stealth with a modular suite built for automation and orchestration:

• Runner H → Executes tasks across Gmail, Slack, Notion, Zapier. Think workflow automation (job hunts, campaigns, travel).

• Surfer H → A real-time web agent for research — open-sourced, 92.2% on WebVoyager.

• Tester H → Generates UI tests from natural language. Built for devs, saves manual QA effort.

All powered by Holo-1, their in-house model — lightweight, cost-efficient, and action-oriented.
Free tier is live. UX feels slightly smoother than Manus. Definitely one to try if you want orchestration-first agents in your browser with a European edge.
→ Explore H Company’s agent stack