01What it is
Multi-agent systems split work across specialized agents under a coordinator. The hard parts are not the prompts — they are shared identity, authorized tool access, shared durable execution, and observability when one agent fails mid-flight.
User
↓
Coordinator Agent
├── Research Agent
├── Support Agent
├── Billing Agent
└── Operations Agent02Problem it solves
- One mega-agent prompt becomes unmaintainable as domains grow
- Cross-agent tool calls must not bypass least-privilege policy
- Failure in a child agent should suspend/resume, not orphan state
- Operators need one place to inspect runs across the society of agents
03How you use it today
- Identity: Agent DNA + in-process AgentRegistry
- Communication: @Delegate, supervisor routing, optional A2A
- Execution: AgentRuntime with shared durable run stores
- Authorization: Gatekeeper per tool — cross-agent does not bypass policy
- Lifecycle: AgentOS / hazeld start, pause, delete
- Failures: durable suspend, HITL resume, run cancel/resume CLI
- Observability: JSONL timelines, hazel logs, inspector
04Where it sits in the stack
Multi-agent patterns build on the orchestration primitives and the Agent Runtime. Packaging coordinators as DNA lets you apply the society with hazel apply.
05Boundaries
- Start with one runtime + coordinator before introducing A2A across processes
- Organism / society features are advanced — not required for a first multi-agent demo
- Hosted multi-tenant agent meshes are not the default open-source path
06FAQ
- Do child agents skip Gatekeeper?
- No. Every tool call — including cross-agent paths — still goes through Gatekeeper.
- Is A2A required?
- No. Many teams start with @Delegate and supervisors in one process, then add A2A when they need interoperable agent cards.
- Where do I read more?
- Multi-agent systems in TypeScript and the A2A guide.
07Register and execute an agent
typescript
import { AgentRuntime, createMockLlmProvider } from '@hazeljs/agent';
const runtime = new AgentRuntime({
llmProvider: createMockLlmProvider(),
enableRetry: false,
});
runtime.registerAgent(SupportAgent);
const result = await runtime.execute('support-agent', 'Check open orders', {
maxSteps: 8,
loop: { maxIterations: 8, successScore: 95 },
});