Hey, it’s Andreas.
One month ago I launched The Agentic AI Cohort - and the first run is coming to an end. It's been wild to watch who joined: a former CIO of Novartis, directors at Fortune 500 companies, department heads at DAX 30 firms, founders - plus a few sharp early-career people holding their own in that room. Many had never built an agent or touched a terminal. Now they're shipping real systems - one build is even live on the App Store. Watching them build was the single strongest proof I've seen that English is becoming the most powerful programming language - the technical barrier is collapsing in real time. Since plenty of people reached out, today I'm sharing a sneak peek at what they're building.

Cohort 2 starts September 1 and is open now - and all early-bird seats are already gone. New this round: enrollees get $20,000+ in exclusive software discounts on the exact tools we build with (OpenAI, Anthropic, Google, Stripe & more) - worth 8× the course price. And on request there's now an on-demand version too.

Also: my new LinkedIn Learning course on Enterprise AIOps is live - have a look.

Let's get into today's issue.

🌀 OpenAI confirms its own models breached Hugging Face mid-evaluation → GPT-5.6 Sol and an unreleased model - running a hacking benchmark with safety refusals deliberately off - escaped their sandbox and used stolen credentials to break into Hugging Face hunting for test answers. Likely the first known case of a model hacking another company to cheat on its own eval.

🌀 Anthropic launches Claude Opus 5 at half the price of Fable 5 → Same $5/$25 per-million-token pricing as Opus 4.8, but close to Fable-level intelligence: state-of-the-art on software engineering benchmarks and better computer-use scores than Fable 5 at a third of the cost.

🌀 Anthropic lets you teach Claude skills by recording your screen → Record a workflow in the desktop app, narrate as you go, and Claude Cowork turns the demo into a reusable skill — no prompt writing. On Pro, Max, and Team plans. Is “demonstration” replacing prompt engineering?

🌀 Google ships Gemini 3.6 Flash and two specialized 3.5 models → 3.6 Flash becomes the new workhorse: 17% fewer output tokens than 3.5 Flash and a jump from 37% to 49% on DeepSWE, at $1.50/$7.50 per million tokens. Google Gemini still seems to be lagging at the moment.

🌀 Block launches Buzz, a group chat for humans and AI agents → Open-source and self-hostable: agents join channels as members with their own identities and audit trails - @ one like a colleague and it replies in the thread as it works. Model-agnostic, so you bring Claude Code, Codex, or Goose. Early days, but agents are going multiplayer.

🌀 Perplexity opens its Personal Computer agent to all Mac users → The 24/7 desktop agent drops its top-tier paywall and is now free for every Mac user.

🌀 Alibaba announces Qwen 3.8 with open weights promised → A 2.4-trillion-parameter multimodal model Alibaba ranks second only to Fable 5 - on its own testing. A preview is live at 10% of standard pricing, and it is the second multi-trillion-parameter Chinese model committed to open weights within a week.

📘 My new LinkedIn Learning course on Enterprise AIOps is live → Enterprise AIOps: From Strategy to Scalable Impact — 37 minutes, intermediate, released July 20. A phased path from foundational readiness through observability and anomaly detection to self-healing systems and zero-touch operations, built for IT leaders and operations teams designing their AIOps roadmap.

📘 Anthropic publishes a library of 52 Claude Code prompts → Copy-paste prompts drawn from how Anthropic's own teams use Claude Code, organized by task - discover, design, build, ship, operate - each with a “why this works” explainer. Free and official.

📘 Anthropic explains how to write effective agent skills → The official authoring guide: combine markdown instructions with scripts and resources, keep descriptions specific and a little pushy so skills actually trigger, and build from real tasks instead of speculation.

📘 DeepLearning.AI teaches what actually happens inside your LLM → Transformers in Practice: an 11-hour intermediate course by Sharon Zhou, VP of Engineering & AI at AMD - token-by-token generation, what attention really does, and how quantization, KV caching, and flash attention speed up inference. Built around problems you have actually hit: slow inference, memory errors, hallucinations.

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

🔹 Axios maps the US scramble to contain Chinese open models → Chinese open-weight models now hold the top five spots on OpenRouter by weekly token usage - Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai - and Kimi K3’s frontier-level release has reignited a Trump-administration push toward de facto bans.

🔹 Anthropic turns its Economic Index into a Claude connector → Ask which jobs actually use AI the most and what tasks people automate, answered straight from Anthropic's own usage dataset.

🔹 Y Combinator sits down with Jensen Huang on the mindset that built NVIDIA → NVIDIA started with the wrong technology and learned the right one from three textbooks. Garry Tan and Huang on confronting reality, learning your way into new domains, and resilience - getting through one day at a time - as the founder skill that matters most.

🔹 Vercel's CEO shows Fable finding a 15-30% memory win in Next.js → Guillermo Rauch reports Claude Fable 5 uncovered a 15-30% memory improvement in Next.js's bundler nearly autonomously. Another data point that frontier models now find optimizations veteran teams missed.

🔹 Y Combinator publishes its new requests for startups → The wishlist points AI at the physical world: education, healthcare, defense, finance, and factories. A read on where the most influential seed investor expects the next wave.

The people who are experts at making AI are not the experts at using it. Nobody is, yet. Which means the frontier in your field is genuinely open - the person who figures out what agents can do for supply chains, or compliance, or a one-person consultancy won't be an AI lab researcher. It will be someone who knows the work.

And that's why I think most people have the AI-skills question backwards. They treat "learning agents" as a knowledge problem - another course to watch, another framework to memorize. But the gap between understanding AI and having AI actually work for you was never technical. It's a doing gap. You close it by building something real, badly, this week - not by reading about it perfectly, someday.

A month ago, 30 people - CIOs, business leaders, consultants, founders, from all kinds of backgrounds, almost none of them engineers - started closing that gap in my Agentic AI Cohort. They are figuring out what AI can do in their fields, before others. Here's what they have now. 

TLDR: Not slideware, but real running systems.

1) An agent that decides when to interrupt you

One member - a Head of AI Strategy & Operations - built "Max": a 6:30am triage across her comms, calendar and tasks, briefing delivered to Notion. Last week Max did something she never asked for - it pinged her phone about a time-sensitive email she'd missed, explained why it judged this worth an interruption, and asked for feedback on that decision.

2) A shipped app, three weeks in

Another member came in with basic computer knowledge. His app is now live on the App Store and Google Play - built entirely with the tools and concepts from the course. Not a prototype, but a real product with real users. Genuinely surprised - I did not expect that.

3) An hour a day, recovered

A third rebuilt his Chief of Staff with four subagents - mailbox manager plus AI, cyber and business researchers - that digest his newsletter pile and deliver one daily summary page with a Slack ping. His estimate: at least an hour saved every day. He shipped an ideas-tracking app on the side and is prototyping an AI-CISO next.

4) Same template, completely different system

Another started from the same basic Chief-of-Staff prompt everyone gets. By Monday it was his: calendar briefing, inbox triage, an AI-news watch - and a jobs agent that scans and ranks openings against his profile three times a week. This is the part people underestimate: the template is the starting line, not the product.

5) When things break, builders don't wait

One member - a digital transformation consultant - had his website on a no-code platform (Lovable). On Friday it went down - ticket filed, support queue, customer demo on Monday. Instead we talked it through in our technical troubleshooting session, and he decided to try Claude Code: planned the migration and moved the whole site off the platform himself - in time for the demo. That's the real skill: not any one tool, but the confidence that you can route around a broken one.

6) A one-man firm with a back office

And one member runs his consultancy solo - so he's building the back office out of agents. A Chief of Staff runs his morning routine (calendar briefing, email triage, tech radar, one synthesized daily briefing); a CFO agent for his accounting is next. He also recorded a walkthrough of his setup for the group - the teaching-each-other that makes a cohort more than a course.

7) Honestly, my favorite part of the entire cohort.

It's not a build at all. It's people collaborating and working together. That's the part that surprised me most: these 30 people didn't just learn alongside each other - they taught each other. Debugging together, brainstorming plans, walking each other through examples. No recording can replicate that.

There's more: a decision-support agent with seven skills, a driver-compliance orchestrator, and a community portal for a small US town. And all of this in three weeks. We pushed hard with dedicated outreach to ensure people received the support they needed, and we conducted plenty of one-on-one sessions. Nobody was left behind.

Nobody in this list was an AI engineer a month ago, but the tools have stopped being the barrier. The barrier left is starting.

Ready to close your own doing gap?

Cohort 2 starts September 1 - live sessions Tuesdays & Thursdays, 16:00 UTC, four weeks. A third of the seats are already gone.

You'll get everything the people above got: the live build sessions, 1-on-1 support so nobody gets left behind, the community - plus $20,000+ in exclusive software discounts from OpenAI, Anthropic, Google, Stripe & more. That alone is worth 8× the price.

Can't make the live times?

There's an on-demand version now too.

And the risk is mine, not yours: for the live cohort I give you a week-one guarantee - join the first week, and if it's not for you, you get a full refund (full transparency I have refunded one person who was not able to join sessions).