Skip to Content

Multi‑Agent Orchestration

The SDK exposes a clean multi‑agent API that composes multiple Agent instances and routes work across them. This hides LangGraph details and lets you think in terms of agents and handoffs.

When to Use

  • You need specialized agents (planner, coder, renderer, safety)
  • You want explicit handoffs and shared state
  • You want a simple supervisor that coordinates multiple agents

Core Types

from isa_agent_sdk import Agent, MultiAgentOrchestrator, ISAAgentOptions

Minimal Example

from isa_agent_sdk import Agent, MultiAgentOrchestrator, ISAAgentOptions
 
planner = Agent("planner", ISAAgentOptions(skills=["documentation"]))
renderer = Agent("renderer", ISAAgentOptions(skills=["documentation"]))
 
# Simple handoff: planner -> renderer -> end
 
def router(state, last_result):
    outputs = state.get("outputs", {})
    if "planner" not in outputs:
        return "planner"
    if "renderer" not in outputs:
        return "renderer"
    return None
 
orchestrator = MultiAgentOrchestrator(
    agents={"planner": planner, "renderer": renderer},
    entry_agent="planner",
    router=router,
)
 
result = await orchestrator.run("Create a 30s storyboard")
print(result.outputs["planner"].text)
print(result.outputs["renderer"].text)

Router Contract

The router receives:

  • state["shared_state"] — mutable shared data
  • state["outputs"] — per‑agent results
  • state["last_agent"] — last agent that ran

It should return the next agent name or None to stop.

Streaming Across Agents

async for msg in orchestrator.stream("Design a creative workflow"):
    agent = msg.metadata.get("agent")
    if msg.is_text:
        print(f"[{agent}] {msg.content}")

When to Use

MultiAgentOrchestrator is best when you have a fixed routing pattern — you know at design time which agent runs after which. The router function gives you full control.

For dynamic handoff where agents decide at runtime when to hand off, or for DAG-based task pipelines across agents, use SwarmOrchestrator.

Notes

  • Each agent uses the standard query() API internally
  • Skills, tools, and execution_mode are configured per agent
  • Use routing + shared state to enforce handoffs and guardrails
  • For dynamic handoff and DAG execution, see Swarm Orchestration
Home