AI Development

A screen in one of my applications told users that a piece of their data was sent to a cloud model. The code deliberately stripped that field before sending. It had been confidently wrong for weeks, and nothing caught it, because nothing tests a sentence. The dangerous artifacts in agent-built software are not the ones that lie. They are the ones with no failing state.

This morning I killed an agent I had forgotten I was running. It had been parked for two weeks, and I only found it because it was holding a name I wanted. Once you are running more than one agent, the interesting engineering stops being prompts and becomes operations. Three things every agent needs before you scale past one.

For eleven days my dashboard reported a broken sitemap on a site whose sitemap was perfectly fine. Cloudflare was challenging my build agent while waving Googlebot straight through. The web is growing a guest list, the free-tier defaults change on September 15, and most internal tooling is on the wrong side of the line.

Eight days after I argued that picking a single best model is the wrong game, Anthropic shipped Claude Opus 5: near-Fable-5 intelligence at half the price, at the same cost as the model it replaces, with a dial that trades effort for money. It did not retire routing. It moved the router inside the model.

A dozen models dropped in about two weeks: Sonnet 5, the GPT-5.6 family, Grok 4.5, Kimi K3, GLM 5.2. If you are still trying to pick the single best one, you are playing the wrong game. The wave did not crown a winner. It retired the question, and made routing the skill that matters.

Marc Andreessen wants to build a sentient sun. Skip the Mars posters and the manifesto hides an operating manual, the same discipline that matters most now that AI made building software nearly free: name the future state, then work back from it.

Fable 5 quietly returned to the Claude Code lineup this week, in the same week the country turns 250. A Fourth of July meditation on the frontier spirit: four frontiers in 250 years, the cognitive one we are standing on now, and why the maker's instinct is a trust, not a trophy.

Anthropic and OpenAI are heading for the public markets on very different balance sheets. The labs might be a bubble. The technology isn't. Here's how to build either way.

Claude now writes more than 80% of Anthropic's own code, and the company just asked the whole industry for a way to hit the brakes. The oldest word for that brake is Sabbath.

Ten days ago the last post here mapped where the labs were planting flags. This week, two of them played opposite moves. Google used I/O to publish an open agent stack, model, IDE, runtime, standard, consumer agent. Anthropic bought the SDK toolchain everyone else used and shut it down. Open vs. moat, in the same week.

Two weeks ago Google Next told everyone this was the agentic era. Then the rest of the industry stopped agreeing in the abstract and started picking specific verticals to own. Anthropic claimed finance. OpenAI counter-launched in security. Microsoft and Notion grabbed the control plane from opposite ends. Here's the new map.

Google made 260 announcements in Las Vegas last week. Most of them are noise. Here's the analyst filter: the four things that actually matter for developers, and an honest look at the product confusion Google still hasn't solved.

Someone built an AI therapist for AI agents. The science behind it is real. The philosophy gets wild. And the marketing strategy is genuinely brilliant. A Christian technologist's take on Delx.ai, AI consciousness claims, and the 2026 frontier of agentic marketing.

Anthropic's next-gen AI model leaked through a misconfigured data cache. The headlines are terrifying. The irony is worse. Here's what actually happened, why the "safety-first" AI company just took a credibility hit, and whether any of this should keep you up at night.

Anthropic shipped more in the first quarter of 2026 than most companies ship in a year. As someone who builds with Claude Code every day, here's what actually changed in my workflow, and what it means for yours.

Amazon's AI coding assistant caused a six-hour outage and millions in lost orders. A new study found 14% of workers are experiencing 'AI brain fry.' The side effects of AI adoption are here, and they're not what anyone expected.

Three tech giants launched competing AI assistants in the same month, all targeting your daily workflow. Here's what Claude Cowork, Microsoft Copilot Cowork, and Google Gemini actually do, and how to think about choosing.

Over 200 internal documents (emails, texts, diary entries, and deposition transcripts) are now public. They tell a very different story than the one OpenAI has been selling.

Google launched Gemini 3, OpenAI shipped real-time coding at 1,000 tokens per second, and Perplexity started pitting AI models against each other. Here's what actually matters.

Anthropic just released Claude Opus 4.6 with agent teams, 1M context windows, and dramatically improved coding capabilities. Here's what matters for developers.