nibzard

My Z.ai setup broke with the latest CC update.

The pitch for the software dark factory is that you swap the dev team for a fleet of agents and software comes out the other end.

You already know the best agent interface ever shipped.

Originally published on the Steel blog.

Originally published on the Steel blog.

Everyone building agents asks the same thing first: which model should I use?

The web is not an API.

AI agents are crawling your site.

I’ve been thinking about what happens when extensibility stops being about loading code and starts being about teaching agents.

I was watching the Clojure documentary this weekend.

The agentic future is seductive.

While building a pile of personal autonomous loops, I kept coming back to the simplest version: Looper.

The interesting thing about Pretext is not text layout.

Skills matter.

Everyone wants the AI-native dev team.

If you want two Claude Code entry points, one for your normal Claude Team or Enterprise login and one for an alternative API provider like z.ai, the cleanest answer is not two installs.

Spending February and March 2025 in San Francisco, going to AI Engineer events in New York, being terminally online on X—it creates a special kind of anxiety.

The licensing fight around chardet is not really about one library. chardet is a small but widely used text-encoding detection library, and the dispute around it is a preview of something larger.

Agents can write code, reason through ambiguity, and call tools.

Here’s a search query that showed up in Google Search Console recently: “browser-use open source agentic ai framework github repository technical documentation showing dependencies, foundation models supported, playwright integration, python libraries, and

I’ve been working with AI coding agents non-stop for over a year now.

Here’s a phenomenon you’ll recognize if you’ve worked with AI coding agents: You ask for a small fix.

We’re living in the era of just-in-time software—tools built and shipped in a single AI coding session.

When you run claude code or use Cursor’s agent mode or any of the growing fleet of AI coding assistants, those agents aren’t just chatting with you.

I checked our onboarding flow for the last seven days.

CLIs are great if you have fingers, patience, and a decent tolerance for “RTFM.”

Software is getting cheaper.

I just watched AI rescue a failed Mistral.ai clone in a single prompt.

AI-native tools win.

Viewed through a Capra-like systems lens: the web is shifting from interface to organism, from clicks to feedback loops.

My site is “static.”

There’s a cliché investors love because it’s true: time kills deals.

X’s recommendation algorithm got a significant rewrite.

I don’t want a vibe-coder.

The assistants don't care where the files live, as long as they're in their expected skills path. # The Drift Problem I’ve been using multiple AI code assistants for a while now—Claude and Codex, each with their own strengths.

Agentic AI isn't a new model capability so much as a new software shape: an LLM inside a loop, with tools, state, and stopping conditions.

If something only works in the UI, the abstraction is broken.

Sorry @UriShaked, my agent did that.

Agent loops make code cheap.

I recently ran an experiment that surprised me.

I’ve made a habit of taking a “sabbatical” every few years to reset and reinvent. 2025 was the year to do it again.

📡 Updated Guide: This article has been updated to reflect the release of GLM-4.7, which introduces interleaved thinking — a new reasoning pattern that interleaves thoughts with actions and responses.

Software is no longer a noun, it’s a verb.

After shipping a bunch of different projects, I’ve seen a pattern emerge: teams focus endlessly on metrics that look good but mean nothing.

# This Isn’t Your Typical Startup Acquisition When I first heard that Anthropic acquired Bun, my initial reaction was probably the same as yours: “Wait, why would a frontier AI lab buy a JavaScript runtime?”

Simon Willison put it perfectly: “You can train a model on a collection of previous prompt injection examples and get to a 99% score at detecting new ones.

The most experienced developers—those who've built systems from scratch, debugged the impossible, and shipped products that millions use—are often the most skeptical about AI coding tools.

I’ve been watching this pattern emerge for months, and it’s finally clicking into place.

Something fundamental has changed in how content gets discovered.

I was staring at another build failure, this time with a particularly frustrating error: [InvalidContentEntryDataError] log → claude data does not match collection schema.