Essays and field notes on building agentic systems — the patterns we use ourselves, the mistakes we'd rather you avoid, and the parts of the craft that don't fit in a demo.
On the same machine, the same model, raw Ollama answers in seconds and Claude Code takes ten minutes. The model is fine. The chrome around it is the variable.
Read the pieceGemma 4 is the first open-weights release of 2026 where the local option is good enough that the question stops being academic. The interesting answer is not that local replaces hosted — it is that runtime is an architectural axis.
The open-source debate keeps polling the maintainer seat. The production-consumer side ships on top of the same libraries every day and is never in the room.
How a 'simple' UI feature led us through five layers of the stack, a library patch, and a fundamental misunderstanding of how streaming actually works.
No list, no drip, no marketing. When we publish something we think you'd care about, we send you a line.