Patterns
Cross-cutting patterns across the book. Each pattern hub links every chapter, Field Note, Recipe, and Lab that uses it.
Side-effecting actions require explicit human approval before execution; non-side-effecting actions don't.
Serverless agent runtimes have a 1-3 second cold start; design around it.
Every framework, tool, or model parameter you add must justify itself with measurable improvement.
When the agent is uncertain or out of its competence range, it escalates to a human or to a different agent — explicitly.
Build the rubric and the gold dataset before you build the agent.
An eval rubric needs categorical failure modes, not just a single score.
One orchestrator agent owns the plan; worker agents have narrow, well-typed jobs.
Observe → Think → Act → Repeat. The simplest possible agent is this loop.
If you can't inspect why the agent did what it did, you can't debug, audit, or improve it.
All tools available to the agent are declared in one schema-validated registry with explicit input/output types.
Every multi-agent step is checked by a deterministic verifier against the original task before the next step runs.
Default to a deterministic workflow before reaching for an agent.