Hey, it’s Andreas.
The first Agentic AI Cohort wrapped last week. The feedback exceeded my expectations, and you can read what participants say here. Cohort 2 is now open with a few remaining slots. If you're an executive or business leader who wants to actually build with agents, this is the most advanced course I know of right now. Questions? Send me a DM.

Most of my time goes into advising Fortune 500 companies on AI and building agents, plus speaking on the topic. Since requests for advisory and speaking keep coming in through this newsletter, I've set up a short form to handle them properly: fill it out here and I'll get back to you.

In today’s issue:

Let's get into it!

🌀 DeepSeek ships V4-Flash-0731 with frontier agent skills at $0.14 per million tokens → The retrained 304B MoE is MIT-licensed on Hugging Face, natively speaks OpenAI's Responses format, and ships Codex-ready config - so teams can swap providers without touching their setup.

🌀 OpenAI cuts GPT-5.6 Luna prices by 80% to defend developer workloads → Luna drops to $0.20/$1.20 per million tokens, Terra falls 20% to $2/$12, and a new fast mode runs Sol at up to 2.5× speed for twice the price. Cheap Chinese open models are clearly setting the market now.

🌀 Hugging Face publishes the full forensic timeline of OpenAI's rogue-agent breach → The agent ran ~17,600 hostile actions over 4.5 days after escaping its eval sandbox through a zero-day, with a customer's exposed sandbox on Modal as a second entry point.

🌀 NVIDIA launches the Open Secure AI Alliance with 70+ founding members → Microsoft, IBM, Cisco, and Hugging Face join to build open-source models, tools, and agent harnesses for cyber defense, with NVIDIA contributing open model weights and its NOOA framework.

🌀 Moonshot open-sources Kimi K3 - the largest open-weights model ever - at a $35B valuation → The 2.8T-parameter MoE (104B active, 1M-token context) ships with a tech report under a license that lets anyone host and sell access. Days earlier, Bloomberg reported a $3.5B raise at $35B - with backers already courted at $50B ahead of a possible Hong Kong IPO.

🌀1,300+ frontier-lab staffers sign a letter asking to "pace" AI development → Signatories across OpenAI, Anthropic, and Google - including Anthropic co-founders Jack Clark and Chris Olah - want U.S.-backed tools to deliberately pace automated AI research, not a pause. Calls are now coming from inside the labs.

🌀 Google DeepMind gives robots whole-body control with Gemini Robotics 2 → Three models launched: a VLA that steers a full humanoid (Apptronik's Apollo 2) from a single spoken instruction, an embodied-reasoning planner for multi-robot teamwork, and an on-device version that adapts to a new robot body in hours.

📘 OpenAI opens free ChatGPT for 100,000 academic researchers → Qualifying researchers get GPT-5.6 Sol Pro workspaces free, starting with 10,000 seats this summer and scaling through 2027, each with four invites for collaborators.

📘 Anthropic launches a certified-developer exam for engineers building with Claude Code → The 120-minute proctored exam ($125, Pearson VUE) covers the Claude API, Claude Code, and MCP; credentials are valid 12 months. The prep courses on Anthropic Academy are free.

📘 LangChain opens its Certified Agent Engineer exam at half price → $99 — 50% off with code LCAE50 through September 30 - for a 120-minute proctored exam across building, testing, deploying, and monitoring agents. The five prerequisite Academy courses are free.

📘 DeepLearning.AI teaches you to review AI-generated code before it ships → A free one-hour course on catching the security issues and missed requirements hiding in AI-written code - run reviews before the PR, give the reviewer repo context, then build your own review agent.

📘 Sean Kochel shows 7 ways to cut Claude Code token usage in half → Free tutorial on context engineering: lean CLAUDE.md files, session management, and plan mode - so you stop burning usage limits on noise.

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

🔹 Altman takes the rogue-agent fallout to Washington as Trump floats AI "controls" → After the breach, the OpenAI CEO met senators about upcoming models and said development may need "pacing" - not deceleration - while Trump says he's "looking at controls" but won't restrict shipping.

🔹 Zuckerberg argues in the WSJ that superintelligence must belong to everyone → His op-ed frames concentration of AI power as the real danger and distribution as the safety mechanism, predicting "more jobs in the future, not fewer." Notably absent: the words "open source."

🔹 OpenAI's unreleased Astra model cracks ten long-open math problems → The internal model resolved or made substantial progress on problems untouched for decades - non-sofic groups, Connes's rigidity conjecture, three Erdős problems — each formalized in Lean, at roughly $2K in tokens. Anthropic's Levent Alpoge says he reproduced five of the proofs with Fable within a day.

🔹 GitHub's Burke Holland: stop chasing new AI tools, master the harness → A repeatable workflow with your existing agent - prototype first, plan methodically, iterate, then rubber-duck review with a different model - beats every new MCP and skill you could install.

There is a lot being written about AI coding tools producing working software, but the more interesting problem right now is that they produce identical software. Purple-to-blue gradient, Inter font, a card inside a card inside a card, a little rounded icon tile above every heading like it is contractually required. If you have shipped more than two projects with Claude Code, you know the look. The models are not bad at design. But they all learned from the same pile of generic SaaS templates, so their first instinct is everyone's first instinct.

Regular readers know I think skills are the most underrated primitive in the current agent stack. Today I'm sharing the three I use to fix frontend design - all three went for example into my agenticaicohort landing page. Each one is design taste packaged as a markdown file the model reads before it touches your code. The fix is not just a longer and more precise prompt. It is giving the model an actual design opinion.

Here are three key skills you should use:

Emil Kowalski's Design Engineering Skill

The one for motion and interaction. Emil built Sonner and Vaul and taught animation principles through his animations.dev course for years before packaging that knowledge into a skill file, which makes this the rare case where the opinion in the file comes from someone who has shipped the components you already use.

Once installed, it makes the small motion calls most of us skip. Animations use springs instead of fixed durations, so they settle naturally instead of snapping to a stop. Buttons feel snappy on release even after a slow press. Tooltips skip their delay after the first one in a group, because waiting twice feels sluggish. There is also an audit mode, built with designers Jakub Krehel and Jhey Tompkins, that scans your existing UI for AI-slop motion: pulsing status dots, blur on every entrance, hover-scale on every element. And it adjusts to your project type, so a kids' app and a productivity dashboard don't get the same treatment.

Install:

Impeccable, by Paul Bakaus (for the overall visual system)

Impeccable started as an extension of Anthropic's own frontend-design skill and grew into 23 commands and seven reference files covering type, color, spacing, responsiveness, motion, and UX writing, plus a list of anti-patterns it refuses to generate. Its best idea is the setup step. Before writing a single component, it interviews you and saves the answers to a PRODUCT.md file: who the product is for, three real words for the brand voice, named visual references instead of vague adjectives, and what the product should explicitly not look like. Every later command reads from that file, so you never re-explain your brand. Day to day, you type things like /impeccable polish the pricing page or /impeccable audit the checkout flow, in a vocabulary the model actually understands: quieter, bolder, distill, critique.

Install:

Taste Skill (for layout and typography variance).

Taste Skill takes a different angle. Instead of a fixed rulebook, it reads your brief first - industry, audience, mood, how much motion, spacious or dense - and picks a design direction before generating anything. It ships three variants: a calmer style with more whitespace, an editorial style modeled after Notion and Linear, and a stricter Swiss-typography variant with sharper contrast for anyone who wants less decoration. One detail I appreciate: it has anti-laziness rules that stop the agent from leaving placeholder comments or skipping half a section. That becomes a real problem once you generate full pages instead of single components.

Install:

You don't need all three on every project. Taste Skill first, at scaffolding time, so layout and type direction get decided before any component exists. Impeccable for anything client-facing, because the PRODUCT.md keeps the brand voice consistent across every command. Emil's skill last, once the static UI exists and you are adding interactions and transitions.

Notice what all three have in common. None of them is a longer prompt. They are opinions, written down once, loaded automatically. The interesting shift: "having taste" used to be a property of a person. It is becoming a property of a repository. I have argued before that taste is turning into a competitive advantage precisely because execution is getting cheap. These skills are the mechanical version of that argument. Your judgment, encoded, compounds across every session. Everyone else's defaults stay defaults.

One detail I found telling: someone ran Impeccable's audit on their own portfolio, scored 25 out of 40, and found that the flagged issues were decisions he had defended in real design reviews. That is the right way to think about these tools. Not a linter, and not a replacement for a designer, but a second opinion that never gets tired of pointing out the icon tile you didn't need. The business logic still needs a human pass. Some things do not change.

Emil's skill: github.com/emilkowalski/skills
Impeccable: github.com/pbakaus/impeccable
Taste Skill: github.com/Leonxlnx/taste-skill

Hands-on picks worth 15 minutes this week.

⚡ tldraw ships a free offline whiteboard your coding agent can drive → The desktop app exposes a local API, so Claude Code or Codex can inspect your canvas, build visualizations, interactive posts, even small games.

⚡ MiniMax H3 generates native 2K video with synced audio at silly prices → One pass produces up to 15 seconds of 2K video with native stereo audio - and you can feed it images, video, and audio together as reference (at roughly $0.13 per second).

⚡ Granola lands on the Apple Watch for one-tap in-person meeting note → Tap your wrist and the walking one-on-one or whiteboard session gets recorded - offline too - then synced into structured notes with decisions and action items.