Blog
These posts are closer to the workbench than the brochure. Agents, MCPs, AI testing, stack discipline, HTMX, data-oriented design, and the practical engineering choices that keep a small team fast without lying to itself.
Recent Posts
These posts cover the engineering decisions behind fast delivery: agents, MCPs, AI testing, stack discipline, internal systems, and the habits that keep software useful after launch.
The pattern is having a moment now, but the idea is older: make the tool useful to humans first, then make it useful to agents. That keeps the interface inspectable, scriptable, and harder to trap inside one protocol.
People blur these together and then wonder why their workflows get weird. Chat is great for exploration. Agents matter when something needs to happen in the world and keep happening without babysitting.
Unit tests are still useful. They are just not enough. Once the system gets probabilistic, the job changes from proving exact text to proving bounded behavior, traceability, and sane failure modes.
AI can generate almost anything. That does not mean your team can review anything. In practice, the stack you can verify is the stack you can safely ship.
Most systems rot because they get modeled around fantasy objects instead of actual data movement. If you start from the transformations, a lot of bad architecture dies early.
No single trick fixes prompt injection. That said, some patterns are still worth using. The goal is not magic. The goal is to make casual failure materially less easy.
Not every internal tool needs to cosplay as a frontend platform. Sometimes the right move is a simpler web app that stays understandable and keeps shipping.
Let the backend describe what is actually possible and let the client adapt, instead of hardcoding fantasy capability into every UI surface.
When the work spans executives, engineers, and operators, writing is not overhead. It is the medium in which the architecture actually becomes usable.
The point of tests in an AI-assisted workflow is not ceremony. The point is that speed without a specification is just the fast production of ambiguity.