The context-engineering paradigm, prompt caching, long-context, lost-in-the-middle, structured outputs, CoT.
When models got capable enough to follow instructions, the bottleneck moved from wording the prompt to curating exactly what the model knows, sees, and remembers at the moment it acts.
The single biggest cost lever for repeated context is a prefix hash that one stray timestamp can silently demolish.
A 1M-token window is an attention budget, not a hard drive — and the model reads the middle of it like a tired commuter skims the middle of a long email.
Two knobs on the decoder — a grammar that constrains the *shape* of every token, and a thinking budget that buys *quality* with latency — and the engineering judgment to know which one the problem actually needs.
A long-running agent is a leaky bucket of tokens — this lesson is how you decide what to keep in the window, what to push to disk, and what to throw away before the bucket overflows.