Google Agent Development Kit (ADK)#

Original | Raw

Summary#

ADK is Google’s open-source, model-agnostic framework for building AI agents. Available in Python, TypeScript, Go, and Java — the broadest language support of any agent framework. Features deterministic workflow agents (Sequential/Parallel/Loop), native A2A protocol for cross-framework interop, and native MCP for tools.

Key Takeaways#

  • Four languages: Python, TypeScript, Go, Java — all first-class. Unique in the space.
  • Model-agnostic: Supports Gemini, Claude, Ollama, vLLM, LiteLLM despite Google origin.
  • Workflow agents: SequentialAgent, ParallelAgent, LoopAgent provide deterministic orchestration without LLM overhead. These are the ADK equivalent of langgraph-agent-orchestration’s graph nodes.
  • Three interaction mechanisms: Shared session state, LLM-driven delegation (transfer_to_agent), and explicit invocation (AgentTool wrapping).
  • Native A2A: First-class Agent-to-Agent protocol support — expose ADK agents as A2A services or consume external A2A agents.
  • Native MCP: Full Model Context Protocol support for tool integration.
  • ADK 2.0 Beta: Adding graph-based workflows, collaborative agents, and dynamic workflows — validating the multi-agent-orchestration graph convergence thesis.
  • Built-in evaluation: Multi-turn eval datasets, local evaluation via CLI/UI. Most frameworks lack this.
  • Deployment: Cloud Run, GKE, Agent Runtime, Ambient Agents.

Architecture#

Hierarchical agent composition:

  • Parent-child via sub_agents (single parent rule)
  • Workflow agents as deterministic orchestrators
  • LlmAgent for reasoning tasks
  • Custom agents (BaseAgent) for specialized logic

Multi-Agent Patterns#

  1. Coordinator/Dispatcher — central router
  2. Sequential Pipeline — ordered steps via state
  3. Parallel Fan-Out/Gather — concurrent execution
  4. Hierarchical Task Decomposition — recursive delegation
  5. Review/Critique (Generator-Critic) — quality loops
  6. Iterative Refinement — LoopAgent with escalation
  7. Human-in-the-Loop — PolicyEngine + SecurityPlugin

Positioning#

DimensionGoogle ADK
Model support10+ (Gemini, Claude, Ollama, vLLM, LiteLLM, etc.)
LanguagesPython, TypeScript, Go, Java
OrchestrationWorkflow agents + LLM transfer + AgentTool
StateSession state + long-term memory service
InteropMCP (tools) + A2A (agents)
Production featuresEval, observability, deployment, streaming
Graph workflowsADK 2.0 beta

See Also#