houmao-agent-loop-pairwise

Agent orchestration converts explicit user intent into a master-owned tree loop plan by normalizing the topology and composing immediate driver-worker edges, ensuring the user agent remains outside the execution loop.

Run control manages a maintained accepted plan by executing start, status, and stop operations through a two-plane control and execution split, keeping the user agent entirely outside the active execution loop.

Supervision loop planning routes mailbox messaging, gateway semantics, and follow-up actions through appropriate sub-skills, while composing driver-worker edges via local-close edge guidance to maintain structured control.

Mermaid control graphs visualize tree loop supervision by mapping immediate driver-worker edges, completion evaluation, and stop evaluation, providing clear graph-structured control visibility for the orchestrated run.

A two-plane control and execution split enforces structured agent run control by separating the normalized master-owned plan topology from execution, preventing ad-hoc orchestration and ambiguous command routing.

Manual-invocation-only tree loop authoring should not be used for implicit intent or multi-plan execution, as it strictly operates exactly one accepted run when the user explicitly requests a single master-owned plan.