Google ADK Guide: Multi-Agent Workflows and Deterministic Nodes

InYourLeague Research · 2026-08-27

Answer in brief

Google ADK composes AI-powered agents and deterministic executable nodes into graph-based, dynamic, collaborative, and template workflows. It also provides experimental runtime routing while helping applications separate responsibilities and limit the data context assigned to each task.

Key facts at a glance

Failure modes and verification

FAQ

What does Google ADK consider a workflow?

Google ADK considers any agent application containing more than one agent or executable node a workflow. A single-agent application is not required to move to that architecture until project needs justify it. See Agents.

How do graph-based and dynamic workflows differ?

A graph-based workflow represents agents, deterministic nodes, and possible decision branches in a flexible execution graph. A dynamic workflow composes the same broad element types through full programmatic code logic. See Workflows.

What is the difference between collaborative and template workflows?

A collaborative workflow gives one agent a dynamic coordinator role over specified sub-agents. A template workflow extends BaseAgent and follows fixed sequential, loop, or parallel execution logic. See Workflows.

Is Agent Routing a standard Google ADK workflow architecture?

The official documentation labels Agent Routing experimental. It uses router functions to select agents at runtime for fallback, A/B testing, or auto-routing. See Workflows.

Does google-adk have an official selectable model ID?

The official source does not publish a selectable model ID for this product. The available documentation describes model integration but does not provide a selectable model ID for google-adk; see Agents.

Sources and freshness

• Official source
• Official source
• Last verified: 2026-08-27

Extended guide

Google ADK directly supports multi-agent applications that combine AI-powered agents with deterministic executable nodes. This google-adk entry was verified against the official documentation on 2026-08-27.

Core composition model

An ADK Agent, also described as an LlmAgent, is a self-contained execution unit. Its basic components are an AI model, task instructions, and an optional set of tools. An agent can perform tasks, interact with users, call external tools, and coordinate with other agents. When an application contains more than one agent or executable node, ADK considers the application a workflow.

ADK does not require every application to adopt a multi-agent or graph architecture. The official guidance presents workflow composition as a way to address increasingly complex instructions, model context-window limits, code modularity, reuse, and the need to interleave non-deterministic model behavior with deterministic code. A workflow can therefore divide responsibilities without forcing one model-driven agent to control every execution step.

Workflow choices

The official workflow overview describes these architectures:

Graph workflows make execution paths and decision branches explicit. Dynamic workflows place path selection in ordinary program logic. Collaborative workflows leave coordination to a designated agent, whereas template workflows apply predefined execution patterns. These options can be selected according to whether a task needs flexible branching, code-controlled behavior, agent-led collaboration, or a fixed and repeatable pattern.

Composition procedure

1. Define each agent from its model, instructions, and optional tools, giving each agent a specific task responsibility.
2. Identify operations that should be deterministic code nodes instead of model-driven decisions.
3. Choose a graph when branches must be represented explicitly, or a dynamic workflow when full programmatic logic should control execution.
4. Use a collaborative workflow when one coordinator agent should work with named sub-agents; use a template for sequence, loop, or parallel execution.
5. Treat Agent Routing as experimental when evaluating runtime fallback, A/B testing, or automatic agent selection.

Data-context boundaries

The workflow documentation identifies limiting data contexts for particular tasks as a structural benefit. Splitting a monolithic agent into agents and executable nodes separates task responsibilities and can keep each task within a narrower relevant context. The documentation does not specify one mandatory context-partitioning scheme, so the precise boundary remains an application design decision.

Verification checklist

• Every agent has an explicit task responsibility.
• Deterministic work is assigned to executable code nodes where appropriate.
• Graph, dynamic, collaborative, or template execution matches the required control pattern.
• Each task receives only its intended data context.
• Experimental Agent Routing is clearly labeled.
• Model metadata does not invent an unsupported identifier.

Model ID status

The official source does not publish a selectable model ID for this product. The documentation explains that agents use an AI model and can integrate models from Google and other providers, but it does not identify a selectable model ID belonging to google-adk itself.

Model availability note: The official source does not specify a selectable model ID.