Notes on building AI systems that hold up in production — context engineering, agent harnesses, loop design and evaluation. Written from what actually broke, not from the launch posts.
An agent is a while-loop around a model. Nobody teaches you how to design that loop, and it is exactly where agents quietly burn money or stall short of done.
The demo takes a weekend and the production system takes a quarter. Here is where that quarter actually goes, and which early shortcuts prove unaffordable.
Vibes-based iteration works until it silently stops working. A real eval suite is the difference between shipping AI features on evidence and shipping them on hope.
The model is a commodity you rent. The harness is what you own: tools, permissions, state and observability. It decides whether your agent is useful or dangerous.
Prompt wording stopped being the bottleneck. What decides whether an LLM feature works in production is which tokens you put in the window, and which you leave out.