Notes on agents, inference systems, and measurement. Written when a result surprised me, or turned out to be measuring the wrong thing.
An inference server stays fast by treating every request as its first. Agents need the opposite. vLLM's agentic layer puts the memory beside the engine, and most of its design follows from one rule about who owns the history.
An agent that grades its own work is a control loop whose sensor sits inside the plant. Every time I fixed my evaluation, it started rewarding a different degenerate agent.
An SGLang result survived every rerun and vanished only when I timed the right interval. Nsight and source tracing revealed when KV-cache pages actually matter.
A generated paper can look finished and be wrong in the two places nobody checks. The useful engineering is the part that fails the build.
A Chrome extension that reads your pulse through the webcam and decides when to interrupt you. The signal took a month. The interruption is a different kind of problem, and I turned it off within a day.
A recommender scores your interest by how long you watch. A browser agent's watch time is however long the model takes to decide. So the agent expresses preferences by thinking.
Song-dynasty supernatural tales were filed for centuries as minor entertainment. They hold the richest geographic and social detail in the corpus, and the reason is structural.