Source: ADK workflows, graph workflows and LoopAgent (ADK docs)

ADK documentation on workflow types: template workflow agents (Sequential, Parallel, Loop), graph-based and dynamic workflows.

Contents
  • What: official documentation pages: Workflows overview, Graph-based workflows, Template workflows, Loop agents
  • Author: Google
  • Date: undated living docs (read 2026-11-14)
  • Original: https://adk.dev/workflows/

Summary

ADK offers four workflow styles: template workflow agents (pre-built Sequential, Parallel and Loop agents that never consult a model for ordering), graph workflows (nodes and edges with conditional routing), dynamic workflows (orchestration in ordinary code) and collaborative workflows (a coordinator agent with sub-agents). Graph, dynamic and collaborative workflows require ADK 2.0. Related pages: graphs, template workflows, loop agents.

Key claims

  • Template workflow agents decide execution order without an AI model, so execution is predictable.
  • Template workflows are marked deprecated in Python, TypeScript, Go and Kotlin docs in favour of graph and dynamic workflows.
  • Graph nodes can be LLM agents, functions/tools or sub-workflows; a node’s output becomes the next node’s input.
  • Conditional edges route on a route value emitted by a node; fan-out and join nodes are supported.
  • LoopAgent stops on max_iterations (Python default None) or when a sub-agent sets escalate; without either it can loop forever.
  • Loop examples pass data via output_key state and {key} placeholders, and rely on an agreed exit phrase, which is brittle.
  • Agent routing (choosing agents at runtime via router functions) is experimental.

Feeds

Suggest changes →