Hey, it’s Andreas.

In today’s issue:

• OpenClaw creator Peter Steinberger joins OpenAI

• A week of viral AI essays - Matt Shumer’s “Something Big Is Happening” and the pushback

• Plenty of new upskilling paths for generative AI and Claude.

• And a full breakdown: the AI vibe coding tool landscape in 2026 - what to use, when, and why.

Let’s get started.

Weekly Field Notes

🧰 Industry Updates

🌀 OpenClaw creator Peter Steinberger joins OpenAI→ OpenAI is doubling down on personal agents by hiring the builder behind the viral open-source assistant, while keeping the project alive as open source under a foundation model. 

🌀 Matt Shumer’s “Something Big is Happening” goes mega-viral → One of the most viral essays on recent technological developments argues that AI is changing the world faster than most people are prepared for. But the counterpoints are compelling as well: Will Manidis’ “Tool Shaped Objects” argues much of today’s AI is performative with limited economic value, and John Coogan notes AI is not COVID - it’s a series of S-curves, not one exponential.

🌀 Anthropic brings Claude Cowork to Windows with full macOS parity → If you use Windows, give it a try as soon as possible.

🌀 ByteDance launches Seedance 2.0 video model and triggers Hollywood IP backlash → Video generation is currently limited to 15 seconds in length.

🌀 Ex-GitHub CEO Thomas Dohmke launches Entire with $60M seed to build agent-human developer infrastructure → Entire’s first product “Checkpoints” logs agent context (prompts, tool calls, files touched, token usage).

🌀 IBM plans to triple entry-level hiring in the U.S. in 2026 → A direct counter-signal to the “AI kills junior roles” narrative: redesign junior jobs around customer-facing work, collaboration, and AI supervision.

🎓 Learning & Upskilling

📘 DeepLearning.AI x Google Cloud x IBM Research launch short course on A2A (Agent2Agent Protocol) → A practical 1.5-hour intermediate course on making agents A2A-compliant (serve + client).

📘 GitHub 10 repos to learn AI agents from real code → A practical path: start with fundamentals, then agent frameworks, then orchestration and then evaluation.

📘 Anthropic launches “Agent Skills with Anthropic” short course for Beginner → Teaches how to build reusable agent “skills” in an open standard, load them on-demand for coding, analysis, and research and then combine them with MCP.

📘 IBM Technology explainer video on OpenRAG and the next step in enterprise RAG → A quick primer on building more modular, portable retrieval pipelines for GenAI.

📘 Prompting tips for Claude Opus 4.6 → A practical guide to writing prompts that steer reasoning, manage long inputs with compaction, and get more reliable multi-step outputs from Opus 4.6.

🌱 Perspectives & Research

🔹 Lex Fridman interviews OpenClaw creator Peter Steinberger in a 3+ hour deep dive on personal agents → Covers how OpenClaw went viral (180k+ GitHub stars), the security risks of skills and agent ecosystems, GPT-5.3 Codex vs Opus 4.6 for agentic coding and acquisition interests.

🔹 Anthropic drops its 2026 Agentic Coding Trends Report (8 predictions) → Core thesis: software work shifts from “writing code” to “orchestrating agent teams” - with longer-running agents, better escalation to humans, broader builder access, and security becoming the main constraint.

🔹 Anthropic CEO Dario Amodei joins Dwarkesh Patel for a wide-ranging discussion → They cover capability timelines, economic impact, and the strategic implications of superintelligence.

🔹 Nader Dabit publishes “You Could’ve Invented OpenClaw” - a from-scratch agent architecture tutorial (~400 lines of Python) → Walks through the core building blocks (sessions, tools, permissions, gateway pattern, compaction, memory, cron, and multi-agent coordination) so you can point an agent at the Markdown and say “build it” to recreate the system.

🔹 IBM and AWS survey 1,000 execs: agentic AI spend is rising faster than readiness → 80% are increasing investment and expect spend to nearly triple by 2027, but most still lack the architecture, data flows, and human-agent KPIs needed to scale.

🔹 Google DeepMind proposes a new framework for “Intelligent AI Delegation”
→ A practical delegation model (task allocation plus authority, responsibility, accountability, and trust) aimed at reducing risk as agents operate in larger delegation networks and the emerging agentic web.

🔹 Microsoft AI chief Mustafa Suleyman interview with FT most white-collar would be automated within 12 to 18 months → He said most white-collar will be automated within 12 to 18 months.

💼  Partner Brief - Infercom (powered by SambaNova)

Run frontier open-source models in Munich, without sovereignty compromises

Most “EU” AI stacks still hide a hard problem: data residency, legal exposure, and operational friction when you scale inference.

Infercom’s new Munich deployment (powered by SambaNova) is built for teams that need performance and compliance at the same time:

• Local hosting + GDPR alignment, designed for EU AI Act requirements 

• No prompt retention: Infercom does not store inference data 

• Sovereign-by-design: processing stays within European borders 

• Up to 10x faster inference than GPU-based systems (LLM-optimized)  LLM-optimized inference for faster speeds than GPU-based systems

• Deploy models like DeepSeek and Llama, or build agents with frameworks like OpenClaw

See available models and get access.

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

In February 2024, exactly one year ago, Andrej Karpathy coined the term "vibe coding" on X.

The idea was simple:
Use LLMs, describe what you want in natural language, accept changes quickly, ship low-stakes projects without obsessing over every line of code.

What started as a meme quickly turned into something much bigger. Over the past year, the development curve has been relentless. Tooling improved. Models got dramatically stronger. Workflows shifted from assistive to agentic. Coding is currently the highest-leverage use case of generative AI.

Not because it replaces engineers.
But because it multiplies them.

On the one-year anniversary, Karpathy himself suggested we may need a better name: agentic engineering.

I prefer that term.

It removes the undertone that this is not “real work.” It captures what is actually happening:

• Agentic because the new default is not that you write 99% of the code yourself. You orchestrate systems that write it. You supervise, steer, correct, and decide with the help of agents.

• Engineering because there is real craft involved. There is structure, judgment, architectural thinking, trade-offs. It is something you can study, practice, and master.

Over the past year, a lot has happened. “Vibe coder” became a real job title and companies started hiring for it.

Corporates picked it up as well:

• Cognizant set a Guiness World Record with 53,199 employees participating in a “Vibe Coding Week”

• Google via Kaggle hosted a 7-day vibe coding competition using Gemini

• Imaguru ran a 48-hour hybrid vibe coding hackathon together with Lovable

Just to name a few.

And if you attend hackathons today, you will notice something else: these tools have become the default way to build software.

More and more companies are integrating AI coding assistants into day-to-day development workflows.

• Anthropic engineers now reportedly create most of new code with Claude

• OpenAI writes now according to own reports 100% of its code with Codex

• IBM is pushing internal AI coding initiatives such as Project Bob and selectively leveraging tools like Claude Code

• Microsoft promotes Copilot commercially, yet many developers inside the company use reportedly Claude Code

• Palantir is leveraging Claude Code for development work

Tech companies are for obvious reasons first movers. But other sectors are following. Spotify says its best developers haven’t written a line of code since December, thanks to AI. Some journalists from CNBC, with no coding experience, vibe-coded a clone of the app Monday, and the company’s stock price promptly crashed.

And the overall numbers show vibe coding is not niche anymore:

• 92% of developers use AI coding tools daily

• 41% of global code is now AI-generated

• 51% of professional developers use AI tools every day

That shift happened in less than a year. But most people lump all existing together and call it “vibe coding.”

That is where the confusion starts.

AI and vibe coding tools are not just one category. They operate differently. They have different levels of autonomy. They serve different purposes. If you want to unlock real leverage, you need to understand which tool to use for which task.

I have noticed significant confusion among people who treat all AI coding systems as interchangeable.

Let’s change that.

There are 5 distinct categories - and they solve different problems.

1. Vibe Coding Apps

2. CLI Agentic Coding Tools

3. IDE Integrations

4. Auto-Debugging Tools

5. Code Review Tools

In this edition, we’ll map the landscape across these five categories:

• What’s currently on the market

• When to use what

• Who should use what (builders, devs, teams, leaders)

• How to combine them into a stack that actually ships

1. Vibe Coding Apps

Vibe coding apps don’t require any setup, allowing you to create a prototype within minutes with very low friction. You can go from prompt to hosted app in no time.

Examples:

• Lovable

• Replit

• Bolt

• Base44

Best for:

• Non-technical builders

• Hackathons

• Internal tools

• Fast MVPs

Of all the tools I tested, Lovable provides the best user experience. Replit is also pretty good, but more advantageous for developers because of its built-in development tools and debugging visibility. All of them offer free tiers, and you can start right away within seconds by clicking the link above.

Use this category when:

• You want speed over control

• You are testing an idea

• You do not care about architecture depth

• You are zero percent technical

CLI Agentic Coding Tools

These tools require a small setup because they are integrated into your CLI (Command Line Interface), which makes them much more powerful than just pure vibe coding apps. They are like full-blown agents, with access to your files, folders, and entire projects. This is where "vibe coding" becomes “agentic engineering”.

Examples:

• Claude Code

• OpenAI Codex

• Gemini CLI

• Warp

They can:

• Refactor large codebases

• Work across multiple files

• Run multi-step tasks

• Operate like junior engineers

With newer model releases, especially Opus 4.6 and GPT 5.3, the capability jump has been dramatic. I consider CLI Agentic Coding Tools one of the most important advancements in AI since GPT-3.5, and if you have not tried them, you can't even begin to understand how powerful they are.

Use this category when:

• You are building something more complex

• You want flexibility

• You work with existing repositories/refactoring

• You need multi-file orchestration

IDE Integrations

Here AI lives inside your IDE, providing real-time help like code suggestions and edits as you work, eliminating the need to switch to separate apps or browsers. This approach became popular after GPT-3.5's release around 2023 but is now declining in favor of more autonomous tools like CLI agents.

Examples:

• Cursor

• Copilot

Strength:

• Real-time multi-file edits

• Inline suggestions

• Visual diff review

Trade-off:

• You are locked into that editor

• Less raw autonomy than CLI agents

Use this category when:

• You are already deep in IDE workflows

• You want augmentation, not full autonomy

• You prefer seeing every change before accepting

Auto-Debugging Tools

The newest category leans heavily into software engineering and is geared towards very technical users. These tools continuously monitor your codebase.

Examples:

• Devin

• Cascade

These tools:

• Review pull requests

• Suggest bug fixes

• Monitor code quality

• Operate continuously

The idea of these tools is more about having a smart coworker working alongside you. They also place a strong emphasis on being used by a typical team.

Use this category when:

• You have scale

• You have multiple engineers

• PR volume is high

• You need consistency

Code Review Tools

Pre-merge AI code review tools represent a niche but a growing segment in developer workflows, primarily used for automated validation before code merges. While not as mainstream as IDE integrations or CLI tools, they provide high-value guardrails for audit-ready processes.

Examples:

• CodeRabbit

• Panto AI

These tools focus on pre-merge validation. They add an extra security and quality layer, which is especially useful if you generate a lot of code with AI. I have not used them extensively and I have heard mixed feedback - some teams love them, others complain about too many false positives. If you work in a regulated industry, or you hold a high bar for code quality, this is still a category worth keeping on your radar.

They:

• Scan for vulnerabilities

• Highlight logic flaws

• Provide actionable PR feedback

• Enforce standards

Use this category when:

• You are scaling a team

• Security matters

• You need audit-ready workflows

So which tools should you use?

The space is moving fast. Half of the tools mentioned here did not exist a year ago. It’s dynamic, and a lot will happen over the next few months.

But one thing is clear: English is increasingly becoming the key programming interface. Syntax becomes secondary. And the people who understand this tools and get hands-on early will lead over the next few years.

Don’t treat the tools as competitors. They’re layers. The leverage comes from combining them deliberately.

My stack looks like this:

• Lovable for ultra-fast prototypes

• Claude Code for heavy lifting and complex projects

• Cursor for visibility and controlled edits (integrated with Claude Code)

Ultimately, you need to find what works for your workflow. My advice: try a few tools and see what clicks. Most have free tiers (just click the links above), and there are plenty of good tutorials.