A meta loop is a layer-3 agentic loop in which an agent or agents control the decisions, state, and plan of internal or task-level processes. Unlike a task loop directed mainly by code, a meta loop reasons about delegation, progress, replanning, and completion. section 9.1 section 9.1.3
Use the subtype that matches how control is distributed:
- Orchestration uses one delegating agent that maintains global state and the plan, calls specialized worker agents as tools, evaluates their results, and decides whether to delegate, replan, or finalize. section 9.3
- Collaboration uses peer agents that share state, contribute, critique, refine one another’s work, and collectively decide when the goal is met. section 9.4
Research orchestration
A research orchestrator keeps an overall strategy, current focus, subtasks, and progress state. It delegates searching to a research worker and analysis to another worker, evaluates the findings, adds or reprioritizes subtasks when gaps appear, and finalizes only when the collective output satisfies the goal. section 9.3
