Skip to content

AI Agent Orchestration

Orchestration is coordinating multiple AI agents (or one agent spawning subagents) to solve work that's too large, too parallel, or too context-heavy for a single agent loop. It's the "multi-agent system" idea applied to LLM agents: a coordinator delegates isolated subtasks and composes the results.

Why it matters

A single agent has one context window and burns tokens on every intermediate step. Orchestration solves two problems: parallelism (research A and B at once) and context hygiene (a subagent does messy work in its own context and returns only a summary). This is exactly the hermes ↔ claude-code relationship on this box.

Patterns

  • Delegation / subagents: a parent spawns children with narrow goals; children can't see the parent's full history, so context stays clean.
  • Orchestrator + workers: one planning agent, several executing agents.
  • Role specialization: e.g. a "VPS architect" agent (claude-code) vs. an orchestrator (hermes) vs. research workers.
  • Guardrails: turn/budget caps, tool allowlists, and verification of subagent self-reports (a claimed success must be checked, not trusted).

Live example on this box

hermes (always-on orchestrator) tasks claude-code (on-box executor) to install/configure services, and dispatches research subagents to build this wiki — without flooding its own context. See stack-sovereignty-box.