Recent content on Niels Berglund
This week is all about trust, and who pays for it. Two posts poke at Google’s Open Knowledge Format from opposite ends: one asks who gets to declare a piece of agent memory “verified”; the other bolts a token pointer onto the frontmatter and hands raw integers between three models. Both land on the same lesson: the format is never the guarantee. 
 Elsewhere, a sharp essay…
This week the agentic era stopped being a demo. It started being an operating model. Steve Yegge shows us a 40-agent city running his 25-year-old MMO; Anthropic makes auto mode the default in Claude Code, with the safety data to justify it; and the MCP-versus-Skills argument finally gets put out of its misery. 
 We’ve also got a .NET agent that turns “generate unit tests”…
A week of plumbing, and I mean that as the highest compliment. MCP shipped its largest revision since launch, tearing sessions out of the protocol core and quietly turning itself into infrastructure the Fortune 500 can actually deploy. Anthropic, meanwhile, deleted over 80% of Claude Code’s system prompt and told us why our carefully curated CLAUDE.md files are probably too long. 
…
Big week in AI land: Anthropic shipped Claude Opus 5, a model that gets uncomfortably close to frontier intelligence at half the price, and published a fascinating look at how their security team defends a codebase where Claude authors 80% of merged code. Meanwhile, “loop engineering” is the new prompt engineering, and yes, there’s a guide for that too. 
 Closer to the metal,…
This week has a bit of everything: distributed systems war stories from Netflix, a rapid-fire catalogue of Claude Code updates you probably missed, and Anthropic explaining how a million lines of Zig became Rust in under two weeks. The common thread? Agents are doing serious engineering work, and the process discipline that makes it possible. 
 But there’s also a cautionary tale that…
This week is bookended by two very different ways of looking inside an AI system: Google quietly publishing a spec for what your agents should know , and Anthropic publishing a genuinely startling paper about what Claude is thinking but never says out loud. In between, two heavyweights of the agent world explain how to work with frontier models and how the systems around them might learn to…
Up until 2021/22, I was a die-hard Windows user. Sure, I’d had brief “flings” with Mac and even Linux over the years (“this year is the year of Linux on the desktop…”; how many times have we heard that?), but I always came back to Windows. 
 Then, in 21/22, I was going to present at SQLBits in London. As I was packing, I tried to boot up my Microsoft…
If there’s a single thread running through this week, it’s that the interesting action in AI has quietly moved from the model’s weights to the text wrapped around them. Skills, prompts, loops, and proxies are where the reliability is won or lost now, and everyone from Nous Research to Microsoft to the Claude Code team seems to have arrived at the same neighbourhood at once. 
…
This week is a study in three different answers to the same question: how do you actually get an AI to do things rather than talk about them? One says the secret is in the architecture, another says it’s in the deployment plumbing, and a third says you’ve had the answer sitting on your laptop for over a year. 
 We’ve got a contrarian deep dive on the GLM 5.2 open-weights…
If last week was about a frontier model launching and getting yanked off the shelves inside 72 hours, this week is the quieter, more useful sequel: how we actually steer, evaluate, and live with the agents we already have. Less fireworks, more plumbing, and the plumbing is where the real money is. 
 We’ve got two fresh Claude Code announcements, a beginner-friendly take on “loop…
What a week to be paying attention. We swing from a Claude Code creator who merges 150 PRs a day from his phone, through the quiet plumbing of agents that run on a schedule, to the launch of the most capable model Anthropic has ever shipped, and then to that same model being yanked off the shelves by a government directive 72 hours later. 
 If you ever needed proof that the AI story in 2026 is…
This week the conversation grew up. Now that agents are reading our files, calling our tools, and acting at machine speed, the interesting questions are no longer “can it?” but “what happens when it remembers the wrong thing, or when someone turns it against us?” So this roundup is a deep dive into the plumbing of trustworthy agents. 
 We start with the seven flavours…
If there were a theme tying this week together, it would be recursion. Anthropic hired the person most associated with using AI to improve AI, shipped a model designed to be honest about its own mistakes, and turned Claude Code into an orchestrator of hundreds of agents at once. The loop is starting to close, and it is fascinating to watch. 
 On the data side, Microsoft Research dropped Data…
This week’s roundup is dominated by a single seismic event: Andrej Karpathy has joined Anthropic. The most beloved AI teacher in the world, a founding member of OpenAI, and former lead of Tesla Autopilot, is now working on using Claude to improve Claude. Meanwhile, Anthropic quietly shipped something that changes the economics of agent development entirely: Claude Managed Agents, a fully…
This week’s roundup is a masterclass in honest reflection. A Microsoft Distinguished Engineer shares hard-won lessons from integrating AI across the full engineering loop on the Aspire team, revealing that the bottleneck doesn’t disappear; it just moves. Meanwhile, Perplexity’s Agents team pulls back the curtain on how they design, refine, and maintain skills at production scale,…
This week’s roundup is a love letter to anyone building with AI agents. A VILA Lab research paper cracks open Claude Code’s source to reveal a 98.4% deterministic harness wrapped around an LLM that gets maximum decision latitude, Anthropic publishes the definitive guide to building skills, and a Claude Code team member argues that HTML is the unreasonable secret weapon for agent…
This week brings a fascinating collision of big ideas and hard-won lessons. Andrej Karpathy lays out his Software 3.0 thesis at Sequoia Ascent, where the context window becomes the new program and “agentic engineering” replaces vibe coding. At the same time, Yann LeCun puts $1.03 billion where his mouth is, betting that the entire LLM scaling paradigm is pointed at the wrong target.…
This week’s roundup spans the full spectrum: from the guts of distributed systems to the soul of AI agent architecture. A Kafka expert one-shots a diskless streaming engine using a formally verified protocol and an AI coding agent. At the same time, a deep source-level analysis reveals that only 1.6% of Claude Code is actual decision logic. Meanwhile, Steve Yegge declares the arrival of the…
This week is all about Anthropic flexing hard. Claude Opus 4.7 drops with stricter instruction following and a new effort level, while Claude Design enters the scene as a full visual collaboration tool. Meanwhile, a 25-person media company reveals how four Notion-powered AI agents replaced their human traffic controller. 
 Beyond the Anthropic orbit, we explore the Agent Harness concept…
What a packed week. Anthropic dominated the headlines with Project Glasswing, a coalition with AWS, Apple, Google, Microsoft, and others using Claude Mythos to find zero-day vulnerabilities in every major operating system and browser. But the quieter stories this week are just as fascinating: Armin Ronacher reports on running his Postgres-native durable execution system in production, someone…
What a week. An accidental (or was it?) leak reveals Anthropic’s most powerful model yet, one they describe as “genuinely dangerous”, while the practical side of AI development keeps maturing at breakneck speed. We’ve got beginner-friendly guides to Claude’s skill system, Anthropic’s own research distilled into a harness engineering playbook, computer use…
This week’s roundup lands at the intersection of two themes that keep resurfacing: making AI agents more reliable in production and more useful for people who aren’t experts in the language or framework they need. From the nuts and bolts of configuring Claude Code’s .claude/ folder, to the emerging discipline of harness engineering, to a delightful story about building a Kafka…
This week, it’s all about AI agents growing up: from programmable skills that turn your coding assistant into a full agent platform, to enterprise security stacks that finally make autonomous agents safe for corporate networks. Whether you’re building skills for Claude Code or trying to convince your security team that AI agents won’t burn down the infrastructure, there’s…
This week’s roundup is packed, from Andrej Karpathy letting AI agents autonomously run ML experiments on a single GPU to Anthropic making bold moves with both the Claude Marketplace and agentic code review. We also look at how coding agents are getting better tooling for context, debugging, and autonomous loops, plus a fantastic engineering case study from Cash App on migrating 450…
This week’s roundup takes us from the depths of Linux kernel lock contention at Netflix to the ambitious vision of federating AI-powered coding agents into a collaborative workforce. Along the way, we explore how data architectures are evolving from passive reporting to autonomous decision-making systems, and how Anthropic is tackling the thorny problem of making sure your agent skills…
This week’s roundup spans the full spectrum of how AI is reshaping software development, from fine-tuning open-source models with a single English sentence to using Claude as your personal coding instructor. We’ve got Block Engineering turning CI failures into actionable fixes with a three-layer intelligence pipeline, Andrej Karpathy declaring that programming fundamentally changed in…
Welcome back to the Building an Event Management System with Claude Code series! In Part 10, we integrated Brevo MCP for email sending, solved the template parameter-mapping challenge with our templates.json registry, and sent our first transactional email through natural language. A milestone, but we ended with three unanswered questions. 
 
 To see all posts in the series, go to:…
This week’s roundup is all about the maturing ecosystem of AI-powered software development and the emerging layers being built on top of it. Anthropic drops a comprehensive trends report predicting that engineers will become agent orchestrators rather than code writers. At the same time, Calvin French-Owen offers a practitioner’s view of juggling Claude Code and Codex in his daily…
Week 07 asks the question the entire industry is wrestling with: if AI can write all the code, what’s actually left for us humans? Armin Ronacher argues we may need entirely new programming languages designed for agents rather than people, while a software engineering retreat and Aaron Stannard both land on the same answer: taste, judgment, and knowing when to stop. 
 The GPT-5.3-Codex…
Welcome back to the Building an Event Management System with Claude Code series! In Part 9, we explored autonomous implementation with Claude Code, letting it implement three import functions with minimal intervention. We learned that autonomous and conversational approaches each have their place, and that verification is always essential. 
 
 To see all posts in the series, go to:…
Week 06 is dominated by Anthropic’s Opus 4.6 launch and its ripple effects: 16 parallel agents autonomously building a 100,000-line C compiler that can compile the Linux kernel, a 1M-token context window that actually works, and Agent Teams turning Claude Code into a parallel development workforce. Meanwhile, Armin Ronacher makes the case for radical minimalism with Pi, a four-tool coding…
Week 05 delivers a fascinating collision between AI’s transformative potential and its sobering risks. While Kieran Klaassen demonstrates how compound engineering with Claude Code enables a two-person team to ship like a team of five (with AI opening 100% of pull requests), and the Model Context Protocol launches interactive UI capabilities that work identically across Claude, ChatGPT, VS…
Welcome back to the Building an Event Management System with Claude Code series! In Part 8, we implemented import_quicket_registrations and import_quicket_checkins , discovered a critical performance issue (20 minutes → 2 seconds!), and teased an interesting question: What happens when we let Claude Code work more autonomously? 
 
 To see all posts in the series, go to:…
This week’s roundup dives deep into the evolving relationship between developers and AI coding assistants, from Armin Ronacher’s thought-provoking “agent psychosis” concept exploring our growing dependence on AI companions to Anthropic’s newly revamped constitution, which treats Claude as an entity requiring context and reasoning rather than rigid rules. 
 We also…
Welcome back to the Building an Event Management System with Claude Code series! In Part 7, we built the foundation of our Import MCP Server, implemented two crucial core functions ( validate_import_data and preview_import ), and successfully connected Claude Desktop to our custom MCP server. We also encountered and fixed (temporarily) a connection issue that we documented as a GitHub issue for a…
Week 03 kicks off 2026 with laser focus on making AI agents production-ready for team environments, from Joe Njenga’s battle-tested playbook for transforming Claude Code into a full-team productivity system to Anthropic’s strategic expansion of Claude Code’s agentic capabilities into everyday work tasks with Cowork. The quality assurance challenge takes centre stage as Njenga…
Welcome back to the Building an Event Management System with Claude Code series! If you’ve been following along, you know we’ve moved from theory to practice, from planning to building. In Part 6, we spent quality time designing our Import MCP Server and uncovering crucial architectural insights that fundamentally changed how we think about MCP servers. 
 
 To see all posts in…
Welcome back to the Building an Event Management System with Claude Code series! If you’ve been following along, you know we’ve made significant progress: we’ve installed Claude Code, integrated it with our IDE, designed an AI-native architecture, set up our cloud database, and created a complete schema with natural language query capabilities. 
 
 To see all posts in the…
Welcome to an unplanned post in the Building an Event Management System with Claude Code series. If you’ve been following along, you know we designed our database schema in Part 5, set up our infrastructure in Part 4, and we’re ready to build import tools in Part 6. 
 
 To see all posts in the series, go to: 

 Building an Event Management System with Claude…
Welcome to the first tech roundup of 2026, where AI agents finally learned to speak UI instead of drowning us in conversational ping-pong, and autonomous coding tools decided sleep is optional. This week’s collection spans Google’s A2UI protocol for civilised agent interfaces, Microsoft’s memory systems that cure AI amnesia, Ralph Loop’s stubborn persistence that ships $50K…
Welcome back to the Building an Event Management System with Claude Code series! In Part 4, we built the infrastructure foundation: a properly structured GitHub repository, a production PostgreSQL database on Neon, and Postgres MCP Pro configured for Claude Desktop. Today, we bring that database to life through conversational schema design and our first natural language queries—using both Claude…
Welcome back to the Building an Event Management System with Claude Code series! If you’ve been following along, you know we’ve covered installation, IDE integration, and architectural planning. Today, we’re moving from planning to building, setting up our PostgreSQL database and having our first AI-powered conversations with it. 
 
 To see all posts in the series, go to:…
Welcome back to Building an Event Management System with Claude Code blog post series! If you’ve been following along, you know we’ve covered how to install Claude Code (which has gotten significantly simpler since we started - see edited 

 Part 1 and 

 Part 2 ) and how to integrate it with your IDE. We even built a basic Node.js application with…
Welcome to the final roundup of 2025! This week brings us an exceptional collection of paradigm-shifting insights as we close out the year: a battle-tested SQL Server log shipping script that’s been running uninterrupted in production for years, a comprehensive deep-dive into router-based agent architectures that are transforming monolithic AI systems, a mathematically rigorous takedown of…
This week explores how AI agents are fundamentally transforming software development, from Hugging Face enabling Claude to orchestrate complete LLM fine-tuning workflows through natural language, to Microsoft’s Agent Lightning adding reinforcement learning capabilities without code rewrites, and Every’s pioneering Compound Engineering methodology, where single developers now accomplish…
Throughout this week, I’ve been diving deep into the philosophical underpinnings of modern programming languages, discovering that Go, Rust, and Zig aren’t just different feature sets, but entirely different worldviews about how we should write software. The AI coding revolution continued its wild ride with Claude Opus 4.5 proving it’s the world’s smartest coding AI while…
This week’s roundup showcases the infrastructure layer that’s transforming AI coding from impressive demos to genuinely trustworthy production tools. Tiger Data open-sourced pg-aiguide to teach LLMs production-ready PostgreSQL, Anthropic released Claude Opus 4.5 with dramatic efficiency gains and frontier-level capabilities, researchers mapped the architectural gulf between shallow and…
This week’s roundup explores the critical shift from vague AI assistants to specialised, instruction-driven agents that actually deliver value, covering everything from GitHub Copilot’s instruction file architecture to sophisticated memory systems that transform digital amnesia into persistent Intelligence. We dive deep into Apache Iceberg’s emergence as the open table format of…
This week’s roundup explores the foundational research papers that transformed AI from an academic curiosity into production-ready technology, covering everything from the revolutionary Transformer architecture to the Model Context Protocol, which enables standardised tool integration. We examine critical performance optimisation techniques for Azure OpenAI deployments, where asynchronous…
This week’s roundup tackles the critical infrastructure challenges emerging as AI agents reshape software development, from Claude Code’s advanced orchestration strategies and the hidden costs of 10x coding velocity to Anthropic’s breakthrough 98% token reduction through code execution with MCP. Microsoft Research’s Magentic Marketplace simulation exposes troubling…