HazelJS LogoHazelJS
HazelJS Blog

AI Agent Orchestration Explained

Author
HazelJS Team
10/21/2026
General
AI Agent Orchestration Explained
← Back to Blog

Orchestration means graphs, supervisors, workflows, retries, and human approval — mapped to real HazelJS primitives without inventing a handoff API.

Agent orchestration is how you coordinate steps, agents, state, retries, timeouts, and human approval so work completes reliably.

In HazelJS, orchestration is not a single class named “Orchestrator.” It is a set of real primitives:

  • AgentGraph and supervisor patterns for multi-step / multi-agent routing
  • @Delegate for agent-to-agent delegation
  • A2A for interoperable agent protocols
  • @hazeljs/​flow for durable WAIT/resume business workflows
  • @hazeljs/​queue when you need BullMQ-style job queues
  • Durable runs + HITL for crash-safe approval gates

There is no first-class API named handoff in @hazeljs/​agent. Prefer the primitives above.

Minimal runtime loop

await runtime.execute('support-agent', goal, {
  loop: { maxIterations: 8, successScore: 95 },
});

The confidence loop stages are plan → execute → critique → validate (optional observe).

Read next