- 2026.05UPD.07
Pick model by task, not vendor.
We run Anthropic and OpenAI side-by-side on every project. Each model has shapes it ships well; capability-per-task ships better products.
- 2026.04UPD.06
Agents need a surface, not just a prompt.
Drasil taught us: an agent without a UI for review is one that gets shut off the first time it surprises someone. Surface design is half the work.
- 2026.03UPD.05
Eval ramps are the build.
Pointzerotwo taught us: most AI work is in the eval loop, not the prompt. Ship the rubric, ship the test set, ship the diffing tool. Prompt comes last.
Products we built for ourselves.
Lab is proof the pipeline works. Same end-to-end shape we run for clients, applied to products we operate ourselves. The point isn't the products. It's the evidence that the pipeline isn't theoretical.
Live products.
Three things we built for ourselves — same pipeline we run for clients.
Recent updates.
Patterns the Lab is shipping or wrestling with right now.