Hey, it's Andreas.
The agent era is starting to split professionals into two groups: people collecting opinions about AI, and people building systems with it.

That is exactly why I built the The Agentic AI Cohort: to help YOU make sure you belong in the second group.

The cohort is almost full, and the room is already very senior: Fortune 500 leaders, consultants, founders, VPs, CEOs, and project leads who have decided to stop watching AI happen and start building with it.

This is not another theory-heavy AI course.

It is live, hands-on, and built around one outcome: you leave with real AI agents running parts of your actual work. No coding required. You direct the AI. It writes the code. You build agents that save time, reduce repetitive work, and give you a practical operating model for where AI is heading next.

Most participants expense it through their company’s L&D budget. Enrollment closes this week. If my work here has helped you think more clearly about AI, this is where we turn that thinking into execution.

Take a look while the last seats are open: The Agentic AI Cohort

Now, on to the news.

In today’s issue:

🌀 Z AI releases GLM-5.2 frontier-class open model → GLM-5.2 brings open weights, a 1M-token context window, strong coding benchmarks, and much lower pricing than closed frontier models. According to most benchmarks this is the new leading open weights model.

🌀 OpenAI wants to certify 300,000 AI consultants → OpenAI is putting $150M behind a partner network with Accenture, Bain, BCG, McKinsey, and PwC.

🌀 SpaceX buys Cursor in $60B all-stock deal → SpaceX is acquiring Cursor to own a core developer workflow, not just another AI tool.

🌀 Microsoft makes Copilot Cowork generally available → Copilot Cowork is now GA worldwide for Microsoft 365 Copilot customers, built for long-running, multi-tool enterprise tasks.

🌀 Midjourney launches medical division with full-body ultrasound scanner
→ Midjourney unveiled a 60-second full-body ultrasound system built with Butterfly Network chips.

🌀 OpenAI and Anthropic poach top Google AI talent → OpenAI hired Gemini co-lead Noam Shazeer, while Anthropic brought in AlphaFold co-creator and Nobel laureate John Jumper.

🌀 Anthropic expects Mythos and Fable to return soon → Anthropic’s Chris Ciauri said the company is “very confident” Mythos and Fable will be available again in the coming days.

📘 Databricks offers free AI and data fundamentals badges → Free on-demand training across Lakehouse, Generative AI, and AI Agent Fundamentals, with shareable Databricks badges for LinkedIn or résumés.

📘 Databricks on what gets agents to production → Sandipan Bhaumik breaks down why a £85K chatbot PoC failed and what fixed it: evals before code, tracing every agent decision, clean data foundations, orchestration patterns, and governance.

📘 DeepLearning.AI launches course on voice for AI agents → A hands-on short course on adding voice to AI apps and agents, from interactive apps to outbound phone calls.

📘 Claude Code setup in 60 minutes → Practical checklist for making Claude Code safer and more useful: permissions, CLAUDE.md, plan mode, hooks, skills, MCP tools, and checkpoints.

📘 Overview Loop Engineering: New agentic AI engineering discipline → Loop engineering is about designing the feedback system around agents: finding work, assigning it, checking it, recording progress, and triggering the next task.

🔹 NVIDIA shows robots improving overnight with agentic research loops
→ Jim Fan’s GEAR lab used Codex agents, real robots, GPUs, and fixed success criteria to improve robotics tasks while the team slept.

🔹 David Sacks pushes back on the Anthropic Mythos ban narrative
→ Sacks argues he did not downplay Mythos’ cyber risk, but the bigger issue is selective enforcement: Anthropic’s models were pulled while comparable OpenAI models reportedly stayed live. The real debate is no longer just model safety - it is whether AI policy becomes technical risk managemen

🔹 Fable 5, Opus 4.8 and Kimi K2.7 David → A side-by-side look at how leading models build the same website. Useful benchmark beyond scores: design quality, layout judgment, and taste are becoming visible differentiators in coding agents.

🔹 OpenAI introduces LifeSciBench for life science research → New benchmark built with 173 biotech and pharma scientists, covering 750 expert tasks across seven research workflows.

🔹 Pew shows AI adoption rising while optimism falls → Pew’s 2026 data shows about half of U.S. adults now use chatbots, but nearly 40% expect AI to make society worse.

🔹 Anthropic shows domain expertise matters most in Claude Code → Anthropic analyzed 400K Claude Code sessions and found users still drive most planning while Claude handles execution. The key signal: agents amplify people who understand the work - coding skill helps, but domain expertise is what raises the ceiling.uary.

My last six months have mostly been this: countless hours inside Claude Code, long conversations with developers shipping real work with it, and a fair number of direct exchanges with people at Anthropic. I'm writing The Definitive Guide to Claude Code for Packt, and I'll keep sharing some lessons here as the book comes together. Here’s one of the most useful ones yet.

For a stretch this spring, Claude Code felt worse. Thinking less, sloppier syntax, fewer tool calls. One popular explanation was that Anthropic had quietly "nerfed" it to save compute. The real story, which Anthropic laid out in an April 23 post-mortem, is a defaults problem. On March 4 they changed Claude Code's default reasoning effort from high to medium to cut the long latency some users were hitting in high mode, a call they later described as the wrong tradeoff. Two other changes stacked on top: a caching bug that wiped the model's reasoning history mid-session, and a verbosity instruction that hurt coding quality. The model didn't get dumber. The cockpit grew more knobs, the default moved without anyone noticing, and almost nobody knew the knobs were theirs to turn.

Below are seven settings you can configure locally to dramatically improve Claude Code's output again. They don't take much time to set up, the changes are small but the impact is big, and I've included the exact commands to set each one.

First, the two ways you change anything. A slash command is typed straight into the Claude Code chat box, the same place you type your requests, starting with /. It takes effect right away, for that session. A settings file is a small text file Claude Code reads on every start, so your choices stick. It lives at ~/.claude/settings.json for your personal defaults across all projects, or .claude/settings.json inside a project for team-wide rules. To open the personal one, paste this into your terminal:

(nano is the simplest built-in editor; use whatever you like.) Everything below tells you which of the two routes to use. The slash command is always the friendly way in).

1. How hard it thinks (effort).

In spring Claude Code started to avoid deep reasoning because of that setting. And the source of most of the "it got worse" feeling. In the chat box:

Newer versions now default back toward higher effort, since one of the things that made Claude Code look so dim this spring was that a lot of work got routed down to medium and low. Anthropic still recommends matching the level to the task rather than pinning everything to high: more effort costs more and runs slower, so it's a deliberate trade of speed for quality. To lock a level in so you don't retype it each session, add it to your settings.json file (the one at ~/.claude/settings.json). And if you'd rather play it safe and keep performance at its ceiling, just set it to high by default. That's what I did.

2. A floor on its thinking.

Since spring, there's a floor on how much compute gets allocated to thinking per turn. When Claude judges a task to be “easy”, it allocates very little on its own. The setting below lets you raise that floor. It's an environment variable in the same settings file, you can adjust:

3. Stop the constant permission pop-ups.
By default, Claude Code asks before nearly every action, so you spend the morning clicking "allow" like a human rubber stamp. My recommendation: type /permissions in the chat box for an interactive menu. The permanent route, in your settings file:

acceptEdits auto-approves file edits but still asks before shell commands. In my personal view it’s a trust decision, not a convenience toggle. Use it in projects you trust, and use "plan" (Claude reads and explores but changes nothing) on code you don't know yet.

You'll also hear about --dangerously-skip-permissions (people call it "YOLO mode"), which skips every prompt entirely. The name is a warning, not a joke. I don't recommend it for general work: it hands the agent your shell with no guardrails, and a single bad command can do real damage.

4. Use the right-sized brain (model routing).
Claude Code comes in different sizes. Opus is the powerful, expensive one; Sonnet is the everyday workhorse; Haiku is the small fast one. Using Opus for a trivial question is like hiring a senior architect to alphabetize a list. Sonnet is good for the basic, everyday tasks. It's the safe default: cheaper, fast enough, and more than capable for the bulk of the work. And to leverage Opus for the genuinely hard problems where the extra horsepower actually justifies its cost. In the chat box:

5. Disconnect the tools you're not using.
Claude Code can plug into outside tools, called MCP connections: a calendar, a database, a service. Each one you leave connected takes up room in Claude's working memory before you type anything, and the idle ones are dead weight. In the chat box:

That shows what's connected and lets you switch off whatever you're not using this sprint. (You'll see dramatic "wastes X tokens" figures online. The exact numbers vary a lot by tool, but the point holds: connected-but-unused is not free.)

The Takeaway

Very little of the spring frustration was the model. It was a default that moved under people who never opened the settings, plus a fog of config folklore where half the variable names are wrong. The fix isn't a secret flag. It's fifteen minutes learning what your own tool actually exposes. You'll feel the difference the same afternoon.