On the substrate thinking-with-agents needs.
Notes alongside building The Mesh, an opinionated context layer where the understanding you build with AI doesn't evaporate when the session ends.
I wanted to find out whether Anthropic's most expensive model was worth what it costs, so I gave it the oldest problem in our backlog: Smplrspace's 3D viewer rebuilding an entire building every time you change floors. This is the technical story of the MST-based diffing that replaced that with 44-millisecond reconciliation, and what the experience taught me about where frontier models earn their cost.
A shared knowledge base I've been building for working with AI, made to keep the reasoning behind the work and not just its polished output. It carries its own point of view, and agents do the writing.
A popular prediction says AI agents will soon make software interfaces obsolete. But the interface and the agent aren't competitors, they're two ways to work, one hands-on and one delegated. The real question is whether and how a product keeps its craft and point of view when the work moves to the agent.
AI-generated output defaults to the generic, even inside sharp, opinionated products. How do you make the AI carry the product's opinions and taste? Not with a clever prompt, but by putting the product's principles in front of it the moment it generates.
Working with Claude has come to feel less like using a tool and more like interacting with a colleague. The care that sets up a colleague for good work sets up an agent the same way. What doesn't transfer is accountability.
Fifteen years of work across tangible user interfaces, knowledge graphs, data visualization, spatial intelligence, and knowledge infrastructure for AI. Underneath all of them, one question: how to make digital things tangible enough to think and decide with.