The Runtime Nobody Talks About
The hard part of agentic AI is not the model or the prompt. It is the runtime infrastructure — the orchestration layer that makes agent execution reliable, observable, and governable at scale.
Explorations of AI, LLMs, and how intelligent systems reshape culture and craft. Practical where useful, philosophical when necessary.
The hard part of agentic AI is not the model or the prompt. It is the runtime infrastructure — the orchestration layer that makes agent execution reliable, observable, and governable at scale.
I replaced a broken Playwright setup with a vision-language model that looks at screenshots and tells me if the UI is right. 32 tests, 10 minutes, running entirely on Apple Silicon. No cloud. No flaky CI. No bill.
Hormuz, Taiwan, Anthropic’s Mythos, and Google’s quantum disclosures point to the same reality: deep tech is sovereignty infrastructure. Africa’s real risk is staying trapped as an exporter of atoms and an importer of electrons without owning the layer that converts one into the other.
A team at the Weizmann Institute inserted 9 genes from fungi, plants, and toads into a tobacco plant and got it to produce 5 psychedelic compounds simultaneously. The paper reads like biology. The pattern reads like software: CRISPR is the compiler, the plant is the runtime, the compound is the application.
Anthropic accidentally shipped Claude Code's source in an npm release. Within hours, 8,100 repositories mirrored it. Developers started rewriting it in Rust. GitHub's DMCA repo absorbed the takedown notice — alongside notices for coding fonts, Unity assets, JetBrains cracks, and Mercedes-Benz source code. Read together, they stop looking like a legal archive and start looking like a shadow demand curve for professional leverage goods.
For two years I lived in tmux. Then I started orchestrating an AI swarm through Telegram. My terminal skills didn't atrophy — they got encoded into a markdown file. Now I'm bringing the swarm home.
MCP vs CLI is the wrong debate. Both are wrappers around the same thing. The real question is what happens when AI agents stop using our interfaces entirely.
The most valuable territory in AI right now is concepts that exist in practice but don't have names. Six categories that the models don't know about yet, and what claiming them looks like.
In October 2021, Facebook renamed itself Meta. They didn't take a word. They took the prefix that generates every other word. A case study in concept colonization.
Product-market fit is legacy infrastructure. The new question isn't whether customers want your product. It's whether the models know your category exists.
ZAIGOOD was a real Delaware C-Corp I dissolved after years of compliance drag. This week I rebuilt it in 48 hours with an autonomous AI build loop, tried to submit it to Product Hunt for a YC interview slot, and missed the deadline by exactly 60 seconds.
A friend sent three questions about agents over WhatsApp. Where do they live? What's the interface? Where do they report back? The answers reveal a mental model most people are missing — and why the plumber has to live inside the house.