google-adk-workflow-agents

name	google-adk-workflow-agents
description	Create ADK workflow agents: LoopAgent, ParallelAgent, SequentialAgent. Use when orchestrating deterministic multi-step workflows — sequential pipelines, parallel fan-out, or iterative loops with exit conditions.

Google ADK — Workflow Agents

Workflow agents are non-LLM orchestrators that control execution flow deterministically (no model calls for routing).

Imports

from google.adk.agents import LoopAgent, ParallelAgent, SequentialAgent
from google.adk.tools import exit_loop

SequentialAgent

Executes sub_agents one after another in order.

from google.adk.agents import Agent, SequentialAgent

research_agent = Agent(
    name="researcher",
    model="gemini-2.5-flash",
    instruction="Research the topic and output findings.",
    output_key="research_output",
)

writer_agent = Agent(
    name="writer",
    model="gemini-2.5-flash",
    instruction="Write a report based on the research output.",
    include_contents="none",
)

pipeline = SequentialAgent(
    name="research_pipeline",
    sub_agents=[research_agent, writer_agent],
)

SequentialAgent Parameters

Parameter	Type	Description
name	str	Unique identifier
sub_agents	list	Agents to execute in order
description	str	For parent agent delegation
before_agent_callback	Callable	Pre-execution hook
after_agent_callback	Callable	Post-execution hook

ParallelAgent

Executes all sub_agents concurrently. Results stored via output_key.

from google.adk.agents import Agent, ParallelAgent

sentiment_agent = Agent(
    name="sentiment_analyzer",
    model="gemini-2.5-flash",
    instruction="Analyze the sentiment of the text.",
    output_key="sentiment_result",
)

keyword_agent = Agent(
    name="keyword_extractor",
    model="gemini-2.5-flash",
    instruction="Extract key topics from the text.",
    output_key="keywords_result",
)

parallel_analysis = ParallelAgent(
    name="parallel_analysis",
    sub_agents=[sentiment_agent, keyword_agent],
)

ParallelAgent Parameters

Parameter	Type	Description
name	str	Unique identifier
sub_agents	list	Agents to execute concurrently
description	str	For parent agent delegation

LoopAgent

Repeats sub_agents until exit_loop is called or max_iterations reached.

from google.adk.agents import Agent, LoopAgent
from google.adk.tools import exit_loop

refine_agent = Agent(
    name="refiner",
    model="gemini-2.5-flash",
    instruction="""Refine the draft. If the quality is satisfactory, call exit_loop.
Otherwise, continue improving.""",
    tools=[exit_loop],
    output_key="draft",
)

review_loop = LoopAgent(
    name="review_loop",
    sub_agents=[refine_agent],
    max_iterations=5,
)

LoopAgent Parameters

Parameter	Type	Description
name	str	Unique identifier
sub_agents	list	Agents to repeat each iteration
max_iterations	int	Safety limit (default: no limit)
description	str	For parent agent delegation

Composing Workflow Agents

Workflow agents can be nested for complex pipelines:

# Research → Parallel Analysis → Loop Refinement → Final Output
full_pipeline = SequentialAgent(
    name="full_pipeline",
    sub_agents=[
        research_agent,
        parallel_analysis,
        review_loop,
        final_output_agent,
    ],
)

Data Flow Between Agents

Use output_key and state to pass data between agents in a sequence:

from google.adk.agents.readonly_context import ReadonlyContext

def writer_instruction(ctx: ReadonlyContext) -> str:
    research = ctx.state.get("research_output", "")
    return f"Write a report based on this research:\n{research}"

writer_agent = Agent(
    name="writer",
    model="gemini-2.5-flash",
    instruction=writer_instruction,
    include_contents="none",
)

Key Patterns

• SequentialAgent: ETL pipelines, multi-step processing, draft→review→publish
• ParallelAgent: Fan-out analysis, multi-perspective evaluation, concurrent API calls
• LoopAgent: Iterative refinement, retry patterns, convergence loops
• exit_loop tool must be given to an agent inside a LoopAgent to break the loop
• output_key stores the agent's text output in state[key] for downstream agents
• include_contents="none" prevents downstream agents from seeing upstream conversation

Related Skills

• google-adk-llm-agent — Agent parameters and configuration
• google-adk-multi-agent — LLM-routed (non-deterministic) multi-agent
• google-adk-function-tool — Creating tools for sub-agents