Field notes · one laptop, several agents The moment you run more than one AI agent on the same code, you have a team. Teams without a pipeline ship whatever happened to be on the workbench. Here is the pipeline, sized for one machine, with the receipts from a single day
I run a personal knowledge-graph pipeline that turns everything I read into a structured artifact: claims extracted, gaps flagged, cross-referenced against everything I've read before. It had quietly been running on the most expensive model available. Default by inertia, not by decision. I switched to a
A vague sense that something should be smarter, and a model gets proposed before anyone has asked what kind of problem this actually is. Some problems are prediction problems, genuinely suited to a model from the start.
Few would ship a compiler that re-parsed your source differently on every run. That is what a prompt is. "The hottest new programming language is English." Karpathy is right, and the line has been carrying more weight than it was built for. Programming languages are compiled or
Deletion has no advocate. AI assistants add code faster than any team could two years ago, removal stays near zero, and the gap becomes migration scope, security triage, and the months a new engineer spends reconstructing intent nobody wrote down. The tilt is measured. ETH Zurich and LogicStar.ai built
The fastest way to build a commodity is to let an agent build it well. The harness is the durable asset: model capability commoditizes, and the evaluation pipeline and organizational context around it do not. I have been building on that premise rather than writing about it. Then I measured
Execution is the typing: turning a decided approach into a working diff. It decays fast. Context fills, compaction runs, and a session that wrote clean code an hour ago starts producing sludge. The fix is cheap: throw it away and open a fresh one. Judgment is the why: the approaches
"When change is easy, the need for it cannot be foreseen; when the need for change is apparent, change has become expensive, difficult, and time-consuming." David Collingridge wrote that in 1980, in a book on controlling technology that leaned on cases like nuclear power and heavy industry.
The code passed every test. It worked in staging. It worked in the demo. It shipped clean. Six hours later it took down checkout, on an input nobody had thought to try, on a path nobody had thought to look at, because the person who wrote it had been moving
Your vibe-code velocity slowing down is not the warning sign everyone says it is. The dominant take treats the plateau as debt coming due: you deferred the architecture, the system can no longer absorb change safely, and every small edit now risks breaking something you no longer understand. Sometimes
Most knowledge tools remember what you read. Almost none of them keep a track record of whether you were right. You can retrieve any fact you have ever filed. You cannot retrieve whether you were right about it. That second gap is the expensive one, and most knowledge systems are
I am a sucker for good words, and one from Tamil literature keeps surfacing the deeper I go into AI: மதக் களிறு, the bull elephant in musth, the periodic testosterone surge that turns a beast of burden into a force of nature. It breaks through fortress gates, shrugs off
Last week Claude Code shipped dynamic workflows. The first thing I pointed one at was the design decision I was least sure of, and verifying it took a prompt instead of a harness. Dynamic workflows are easy to underrate. Claude Code can now write its own multi-agent harness on
Anthropic's CEO used part of his own developer conference to explain why writing the code is no longer the hard part. His framing was Amdahl's Law: when you speed code generation up several times over, the work you did not speed up becomes the constraint. Review,
When AI was a private tool, PMs wrote for two audiences: business stakeholders and developers. Deck and PRD. That world is closing. PMs now write for three — and the third audience fails in a way the first two never did, in a way that's invisible at the moment
AIStrategy
Our position for months has been that the harness around a model, its scaffold, prompts, and configuration, does more work than its weights on whether the product holds up in production. The AI market is still mispricing this. Procurement and investment decisions anchor on model capability: benchmarks, parameter counts, training
AgenticAI
A frontier AI model read nine labeled debugging traces — three from itself, three from each of two peer models, all on the same bug. In a single API call it wrote 101 words describing what the peers did that it didn't. Those 101 words, pasted verbatim into its
AgenticAI
AI agents can already think overnight. They monitor research, evaluate evidence, run scheduled analyses. The compute is async. The delivery isn't. When an agent discovers something at 3am, it has no native way to reach you in the medium where you'd actually act on it. The
AIStrategy
In March 2026, British dairy farmers were pouring milk down the drain. That same month, memory chip stocks lost $100 billion in market value. One is an agricultural commodity crisis. The other is a technology repricing driven by a Google Research paper. They share no supply chain, no customer base,
AIStrategy
Roman Mars mentioned wanting to do a 99% Invisible episode on the Brannock Device. He was looking for a design story. I think it's a strategy one. The Brannock Device is that metal foot-measuring tool bolted to the floor of every shoe store. One company in Syracuse,
AIResearch
Last Tuesday I opened my inbox to 247 unread items tagged "AI." Newsletters, RSS feeds, research drops, podcast transcripts, analyst notes. By Wednesday morning there would be another 200. I have been tracking this number for months. It does not go down. If anything, the production rate of
AgenticAI
Coding agents hallucinate APIs and forget what they learn between sessions. Andrew Ng just built infrastructure to fix it. Context Hub is an open-source tool that gives coding agents curated, versioned documentation instead of relying on training data. That premise, straight from the README, describes the exact failure mode
AIEvaluation
Anthropic publishes skills for Claude Code: structured task definitions that tell the model what to build and how. They're designed for Claude. I ran them on GPT-5.4 and Gemini 3.1 Pro too. Claude won. Claude Opus 4.6 averaged 0.825 across four Anthropic skills,
AgenticAI
Most companies treating agent-pickability as a distribution problem are solving the wrong layer. Getting into training data, appearing in MCP registries, reducing integration friction to five lines of code: that's necessary. It's also table stakes within eighteen months. The durable advantage is further down: whether