LLM Engineering

LLM engineering is where I file the essays that are about a specific model capability or protocol change and what it does to a system built before it existed. Extended thinking arrived and the invoice doubled because every easy query bought a reasoning path it never used. Million-token windows arrived and quality across the window turned out to be uneven. A new model dropped and someone bumped the string without running an eval. The MCP spec went stateless and most servers did not.

These are not failures of the models. They are the normal cost of building on a surface that moves every few weeks, and the essays are about keeping your own system stable while it does.

Start with reasoning effort budgeting, which is the clearest example of a capability that should be a per-query decision rather than a global switch. Then the model upgrade migration gate for a repeatable way to adopt a new model, and the long context essay for measuring effective context instead of trusting the advertised number.