NOTES / FRAMEWORKS AND SDKS

Google ADK

Google's Agent Development Kit: LLM agents plus Sequential, Parallel and Loop workflow agents, composed into hierarchies, across Python, TypeScript, Go and Java.

UPDATED 18 NOV 2026#google#adk#multi-agent#workflow-agents

What it is

The Agent Development Kit (ADK) is Google’s open-source framework for building and deploying agents. Its docs describe an agent, or LlmAgent, as a self-contained execution unit designed to act autonomously toward specific goals, with support in Python, TypeScript, Go and Java1. Google’s launch post stresses a multi-agent design: modular applications built by composing specialised agents in a hierarchy, with coordination and delegation3.

Agent types

Diagram: an LlmAgent that decides at run time, and SequentialAgent, ParallelAgent and LoopAgent that run sub-agents in a fixed pattern

Diagram: my own.

  • LlmAgent uses a model to reason, choose tools and decide what to do.
  • Workflow agents (SequentialAgent, ParallelAgent, LoopAgent) manage the execution flow of their sub-agents in defined patterns, for predictable pipelines4. They do not need a model to decide the order2.
  • Custom agents extend a base agent class for non-LLM logic.

This mirrors the split in Workflows versus agents: fixed control flow where you know the path, model-chosen flow where you do not. Sequential, parallel and loop correspond to prompt chaining, parallelisation and an evaluator loop.

Composing agents

ADK docs define a multi-agent system as an application where different agents, often in a hierarchy, collaborate or coordinate toward a larger goal2. The primitives include agent hierarchy through sub-agents, workflow agents, and interaction mechanisms. Control can pass by LLM-driven delegation, where an agent transfers to a sub-agent4, or by wrapping an agent as a tool the parent calls, which keeps the parent in control.

The docs also say any application with more than one agent or executable node counts as a workflow, and that ADK does not force you to move from a single agent to multi-agent or graph-based designs1. That matches good practice: start with one agent (Choosing a pattern).

Strengths and trade-offs

My assessment from the docs:

  • Strengths: clear vocabulary for deterministic versus model-driven control, hierarchies of specialists, several language runtimes, and a path to Google Cloud deployment.
  • Watch for: the framework is closely aligned with Gemini and Google Cloud, though it supports other models. Check model support for your provider.
  • Fit: teams on Google Cloud, or those who want explicit workflow agents as the backbone with model-driven agents inside.

Practical tips

  1. Begin with a single LlmAgent and tools. Add workflow agents when the process has a fixed shape.
  2. Use shared session state to pass data between steps, and name outputs explicitly.
  3. Write clear agent descriptions, since delegation depends on them.
  4. Bound loops with a maximum iteration count (Stopping conditions).
  5. Evaluate trajectories as well as final answers (Why evals).

For interoperability across vendors, see A2A. Compare options in Choosing a framework.

Sources

  1. [1]Agents — Google, Agent Development Kit docsPARAPHRASED
  2. [2]Multi-agent systems in ADK — Google, Agent Development Kit docsPARAPHRASED
  3. [3]Agent Development Kit: Making it easy to build multi-agent applications — Google Developers Blog · 2025PARAPHRASED
  4. [4]ADK documentation home — GooglePARAPHRASED

© Ashish Kumar. All rights reserved. Quoted and cited material belongs to its owners. How I attribute sources.

Product, company and framework names are trademarks of their respective owners. Mentioning them does not imply affiliation or endorsement.