Every AI invocation is a typed loop.

Most agent frameworks let you wire up LLM calls, tools, and orchestration logic in code. That works for prototypes. It breaks down when you need to trace a production failure through five nested calls, swap the model for one step without breaking the rest, or compose loops written by someone else without rewriting them to fit your framework's idioms.

LSEM (Loop Stack Execution Model) takes a different approach. It identifies a single primitive — the typed loop — and treats it as first-class. Every prompt, every tool call, every RAG lookup, every agent handoff is a loop with a defined input schema, a defined output schema, a lifecycle, and a tracing surface. Composition, observability, and reliability become runtime concerns rather than glue-code concerns.

LSEM is intentionally small. The spec defines the primitive and its lifecycle. It does not define a coordination protocol, a deployment model, or a user interface — those belong to layers above. The reference runtime, LSR, makes opinionated choices about each.