What I'm reading, building, and thinking about. Mostly agentic AI.
How the writing is organised. Branches with nothing in them yet are shown greyed out — that's the shape I'm working towards, not a claim that it's filled.
How the work actually gets done: harnesses, skills, tool boundaries, and where the human stays in the loop.
Treating agent systems as graphs — what the nodes are, what the edges carry, and which shape survives contact with a real task.
Nodes are agents or steps, edges are control flow. Fan-out, fan-in, handoffs, supervisors, and stigmergic alternatives to all of it.
Entities and relations as the retrieval substrate. Traversal instead of similarity, and when that trade is worth making.
What each node in a system is allowed to see. Assembly, compaction, and the shape of what gets handed down.
Talking to software instead of typing at it — latency, turn-taking, and why voice agents are harder than the demos suggest.
The evergreen explainers. What agentic AI is, why agency is the shift that matters, and how much of it to hand over.
Two Anthropic ideas, a year apart, that keep getting conflated. What each one actually solves, Anthropic's own answer on how they relate, and a two-question test for picking between them.
Google published a framework for reasoning about AI-assisted development stakes. I ran it against my own harness — here's what held up, and where it didn't.
What ant colonies taught me about designing multi-agent knowledge systems — using this project's own over-built first attempt as the cautionary tale.
A way of thinking about how much agency we hand to AI systems, from recommendations up to autonomous execution with oversight.
What if AI could learn your methodology once, then apply it automatically? Claude Skills as a shift from code as reusable unit to expertise as reusable unit.
How I explain AI systems that can actually take action and make decisions — not just analyze data or generate content.
Why the real game-changer isn't smarter AI — it's AI that can actually do things.
Speculation on what AI with agency unlocks, and what I think is coming.