Skip to content

Foundations

Foundations

The vocabulary and mental models the rest of the course leans on. Read these once and the six parts of a loop stop being jargon.

The vocabulary and mental models the rest of the course leans on. Read these once and the six parts of a loop stop being jargon.

Loop engineering is the practice of designing the system that prompts an AI agent — its trigger, instructions, guardrails, verification, state, and logging — rather than prompting it by hand. This section defines that system before you build one.

PageWhat it covers
ConceptsThe core ideas, in the order they build on each other
GlossaryEvery term the course uses, defined once
The four layersHarness, loop, agent, and model — and what lives where
Mental modelsHow to reason about a loop you cannot watch run
PrimitivesThe building blocks a loop composes
Primitives matrixWhich primitive to reach for, side by side

The six parts of a loop

Every production loop declares all six. A loop missing one of these is the loop that surprises you at 3am.

  • Heartbeat — the trigger that starts a beat: in-session, conditional, scheduled, or event-driven.
  • Body — the capabilities it may use: files, tools, skills, connectors (MCP).
  • Spine — durable state between runs. With a spine an interrupted loop resumes; without one it restarts.
  • Stopping condition — a machine-checkable definition of done: success, limit, or no-progress.
  • Checker — the separate reviewer of the output. A loop never grades its own work.
  • Human gate — the placed decision a person must make before anything ships.

Next: Part 1 — The Shift · Back to the course.