Long-form notes from the field.

There is no list price for an AI agent, and anyone who gives you one before understanding your systems is guessing. The cost is set by five drivers — and most of them keep spending after launch.

A working agent demo takes an afternoon. An agent you can put your name on in production takes weeks. The gap between the two is the whole job.

The build-vs-buy question is framed as two doors. There is a third, and it is the one that leaves you owning the asset instead of renting your own workflow back from a vendor.

A working blueprint for an AI-native SaaS on Claude — the Messages API at the core, typed tool use for the work, and the multi-tenant, billing, and eval scaffolding that keeps it shippable.

The model leaderboard is the worst place to start a procurement decision. Here is the framework we actually use when a client asks whether to build on Claude or GPT.

The Model Context Protocol is an open standard for connecting AI models to the tools and data they need. One interface, written once, that any compliant model can use.

Two techniques get conflated constantly. They solve different problems, fail in different ways, and cost different amounts to keep alive. Here is how we decide.

Sovereignty is not a feature you buy — it is a set of decisions about where data lives, who controls the model, and which regulator you answer to. This is how to weigh them before you sign anything.