Multi-Agent Orchestration — Google ADK + Temporal

This recipe builds a multi-agent dispatch pipeline using Google ADK and the Google ADK integration for Temporal. Two agents reason in parallel, a third synthesizes their output and submits a structured decision — and every LLM call and every I/O tool call is a durable Temporal Activity. (The final decision write is a plain in-workflow tool — see Structured output via session state.)

ADK ↔ Temporal mapping​

Pipeline​

                      ┌──────────────┐
                      │ Order input  │
                      └──────┬───────┘
                             │
            ┌────────────────┴────────────────┐
            │       ParallelAgent             │
            │ ┌─────────────┐ ┌─────────────┐ │
            │ │ Fleet Agent │ │ Customer    │ │
            │ │             │ │ Agent       │ │
            │ └─────────────┘ └─────────────┘ │
            └────────────────┬────────────────┘
                             │
                      ┌──────▼───────┐
                      │ Dispatch     │
                      │ Agent        │
                      └──────┬───────┘
                             │
                  tool_submit_assignment
                             │
                      ┌──────▼───────┐
                      │ Workflow     │
                      │ returns      │
                      │ AssignmentOut│
                      └──────────────┘

• Fleet Agent — checks driver positions, capacity, ETAs.
• Customer Agent — checks order priority, deadline, urgency. (Runs in parallel with Fleet Agent.)
• Dispatch Agent — reads both assessments from session state, picks a driver, submits the structured decision via tool_submit_assignment.

This recipe highlights:

• Multi-agent composition with ADK — ParallelAgent + SequentialAgent driven by a single Runner invocation.
• LLM-as-activity — TemporalModel routes every LLM call through an invoke_model activity. Each call appears as a separate event in workflow history with retries, timeouts, and a Temporal-UI summary.
• Tool-as-activity — activity_tool wraps a Temporal Activity so the agent can call it. Each of these tool invocations is its own activity event. (The Dispatch agent's final tool_submit_assignment call is a plain in-workflow tool, not an activity — it only writes to local session state.)
• Structured output via session state — the final agent calls a Python tool that writes to tool_context.state. The workflow reads that key back after the runner completes.
• Sandbox-safe ADK imports — ADK and google.genai are imported under workflow.unsafe.imports_passed_through() so the workflow sandbox does not reject them.

Prerequisites​

• Python 3.10 to 3.13
• uv for dependency management
• A running Temporal Dev Server (temporal server start-dev)
• A Google API key with access to Gemini (GOOGLE_API_KEY)

Setup​

uv sync
export GOOGLE_API_KEY='your-api-key-here'

Running​

In one terminal, start the Temporal Dev Server:

temporal server start-dev

In a second terminal, start the worker:

uv run python worker.py

In a third terminal, kick off an assignment:

uv run python start_workflow.py

You should see a final assignment printed, e.g.:

Assigned driver: driver-a
Reasoning:       Closest available driver with capacity for VIP order

Open the Temporal UI at http://localhost:8233 to see each LLM call and tool call recorded as its own activity in the workflow history.

Architecture​

multi_agent_adk_python/
├── activities/
│   └── tools.py             # @activity.defn — fleet status, priorities, route info
├── models/
│   └── models.py            # AssignmentInput / AssignmentOutput (pydantic)
├── workflows/
│   ├── assignment_workflow.py  # agents + workflow inline
│   └── _activity_tool.py       # ADK ↔ Temporal tool adapter
├── worker.py
├── start_workflow.py
└── tests/
    ├── test_activities.py
    └── test_workflow.py        # workflow logic with a mocked invoke_model

Activities as agent tools​

Each tool the agents can call is a Temporal Activity:

File: activities/tools.py

async def tool_get_fleet_status() -> str:
    """Return current fleet state: driver positions, capacity, and status."""
    return (
        "Fleet status:\n"
        "- driver-a: pos=(36.1147, -115.1728)  capacity=2/3  status=AVAILABLE\n"
        "- driver-b: pos=(36.1099, -115.1750)  capacity=0/3  status=AVAILABLE\n"
        "- driver-c: pos=(36.1162, -115.1745)  capacity=3/3  status=FULL\n"
        "- driver-d: pos=(36.1213, -115.1700)  capacity=1/3  status=AVAILABLE\n"
        "- driver-e: pos=(36.1080, -115.1760)  capacity=2/3  status=DISCONNECTED"
    )

In a real system the body would query a fleet database or hit an internal service. Here we return canned strings so the recipe runs without any backing infrastructure.

TemporalModel — every LLM call is an activity​

TemporalModel (from temporalio.contrib.google_adk_agents) is an ADK BaseLlm whose generate_content_async runs through a Temporal Activity. The plugin registers that activity (invoke_model) on your worker for you.

File: workflows/assignment_workflow.py

def _fleet_agent() -> Agent:
    return Agent(
        name="fleet_agent",
        model=TemporalModel(
            DEFAULT_MODEL,
            activity_config=ActivityConfig(
                task_queue=TASK_QUEUE,
                summary="Fleet Agent — LLM reasoning",
            ),
        ),
        ...
        tools=[_fleet_status_tool, _route_info_tool],
        output_key="fleet_assessment",
    )

activity_tool — every tool call is an activity​

activity_tool wraps a @activity.defn so it presents to ADK as a regular Python tool, but the call body executes via workflow.execute_activity:

File: workflows/assignment_workflow.py

_fleet_status_tool = activity_tool(
    tool_get_fleet_status,
    task_queue=TASK_QUEUE,
    summary="Fleet Agent — get fleet status",
    start_to_close_timeout=timedelta(seconds=10),
    retry_policy=_TOOL_RETRY,
)

The local workflows/_activity_tool.py adds graceful failure on top of the upstream temporalio.contrib.google_adk_agents.workflow.activity_tool: when an activity execution fails (retry policy exhausted, non-retryable application error, timeout), the wrapper catches the ActivityError and returns it to the LLM as a string so the agent can adapt instead of crashing the pipeline. (The retry attempts still appear in workflow history.) Programming bugs — e.g. argument-binding errors — are not caught and propagate normally. Upstream activity_tool (temporalio>=1.25) already handles multi-arg activities and local non-workflow ADK runs.

Composing the pipeline​

File: workflows/assignment_workflow.py

def build_assignment_pipeline() -> SequentialAgent:
    """Compose the full pipeline: Parallel(Fleet, Customer) → Dispatch."""
    return SequentialAgent(
        name="order_assignment",
        sub_agents=[
            ParallelAgent(
                name="assessment_parallel",
                sub_agents=[_fleet_agent(), _customer_agent()],
            ),
            _dispatch_agent(),
        ],
    )

Fleet and Customer agents run concurrently inside ParallelAgent. When both finish, the Dispatch agent runs. Each sub-agent's output_key (fleet_assessment, customer_assessment) writes its final response into session state — that's how the Dispatch agent gets the upstream context without passing it explicitly.

Structured output via session state​

The final agent calls a plain Python tool (not a Temporal activity) that writes the decision into ADK session state:

File: workflows/assignment_workflow.py

async def tool_submit_assignment(
    tool_context: ToolContext,
    driver_id: str,
    reasoning_summary: str,
) -> str:
    """Submit the final order assignment. You MUST call this tool with your decision.

    Args:
        driver_id: The driver to assign the order to (e.g. "driver-a").
        reasoning_summary: One-sentence explanation of the choice.
    """
    tool_context.state["assignment"] = {
        "driver_id": driver_id,
        "reasoning_summary": reasoning_summary,
    }
    return "Assignment submitted."

The workflow runs the pipeline to exhaustion, then reads that key back:

File: workflows/assignment_workflow.py

async for _ in runner.run_async(
    user_id="workflow",
    session_id=session.id,
    new_message=Content(parts=[Part(text=prompt)]),
):
    pass

updated = await session_service.get_session(
    app_name=APP_NAME,
    user_id="workflow",
    session_id=session.id,
)
state = (updated.state if updated else None) or {}
assignment = state.get("assignment") or {}

return AssignmentOutput(
    driver_id=assignment.get("driver_id", ""),
    reasoning_summary=assignment.get(
        "reasoning_summary", "Dispatch agent did not submit a decision."
    ),
)

This pattern — a tool call that writes structured output into session state — is how you reliably extract a typed decision from a multi-agent pipeline that emits many intermediate events.

Sandbox-safe ADK imports​

ADK and google.genai aren't safe under Temporal's workflow sandbox by default, so they're imported under workflow.unsafe.imports_passed_through():

File: workflows/assignment_workflow.py

with workflow.unsafe.imports_passed_through():
    from google.adk.agents import Agent, ParallelAgent, SequentialAgent
    from google.adk.runners import Runner
    from google.adk.sessions import InMemorySessionService
    from google.adk.tools import ToolContext
    from google.genai.types import Content, Part
    from temporalio.contrib.google_adk_agents import TemporalModel

The GoogleAdkPlugin registered on the worker handles the rest of the sandbox passthroughs and deterministic-runtime overrides ADK needs (UUIDs, clocks).

Worker — one queue, one plugin​

File: worker.py

client = await Client.connect(
    "localhost:7233",
    data_converter=pydantic_data_converter,
)

# GoogleAdkPlugin registers the `invoke_model` activity (used by
# TemporalModel for LLM calls) and provides the workflow-sandbox
# passthroughs and deterministic runtime overrides ADK needs.
worker = Worker(
    client,
    task_queue=TASK_QUEUE,
    workflows=[MultiAgentAssignmentWorkflow],
    activities=[
        tool_get_fleet_status,
        tool_get_order_priorities,
        tool_get_route_info,
    ],
    plugins=[GoogleAdkPlugin()],
)

await worker.run()

GoogleAdkPlugin registers the invoke_model activity that TemporalModel routes LLM calls to — you don't need to register it yourself.

Extensions​

This pipeline is the minimal multi-agent shape. Natural ways to extend it:

• Real backing services — replace the canned tool bodies with real database queries, route APIs, or internal microservices.
• More parallel branches — add agents to the ParallelAgent (an inventory agent, a credit-check agent, a fraud-screen agent) — each posts its assessment via output_key for the synthesizer to read.
• Multiple sequential stages — add a validation or post-processing stage after Dispatch by extending the outer SequentialAgent.
• Human-in-the-loop on the synthesizer — pause before the final decision via a Temporal Signal. See the human_in_the_loop_python recipe.
• Graceful degradation — fail-fast retry policies on a sub-agent's tools so the synthesizer can decide with partial data when an upstream service is unavailable.