Loop Orchestrator

Overview

The agent lifecycle is six phases. Each phase is a loop: enter with a goal, run the inner work, and exit only when a deterministic condition is met. The loop, not the agent's confidence, decides when a phase is done.

DEFINE → PLAN → BUILD → VERIFY → REVIEW → SHIP
  │       │       │        │        │       │
  loop    loop    inner    checker  loop    loop
                  loop

Two invariants govern everything:

1. Termination is enforced by criteria, never by the agent's claim of done. Every loop has an exit condition you can check with evidence (a file exists, tests pass, no blocking issues). If the exit condition is unmet, the loop keeps going or stops for a human — it never "completes" on assertion.
2. The entity that verifies is separate from the entity that produces. The BUILD phase does not grade its own work. VERIFY and REVIEW run as independent checks with fresh eyes.

When to Use