Google’s Python agent kit adds deterministic graphs to LLM apps

What it does

ADK 2.0 is Google’s open-source Python toolkit for building agentic applications. It wraps LLM calls into code-first Agent primitives and orchestrates them through a graph-based Workflow runtime that handles routing, state, and retries. You define agents and their handoffs in Python, then test them through a bundled local CLI and web UI.

The interesting bit

Version 2.0 replaces the vague promise of “sophisticated agents” with concrete control-flow primitives: a graph execution engine and a Task API for structured agent-to-agent delegation, including nested workflows and human-in-the-loop checkpoints. The underlying message is that production multi-agent systems need deterministic orchestration, not just better prompts.

Key highlights

• Graph-based Workflow runtime with support for loops, fan-out/fan-in, dynamic nodes, and nested workflows.
• Task API for structured delegation between agents, covering multi-turn handoffs, single-turn output control, and mixed patterns.
• Bundled local CLI and web UI for testing multi-agent directories without extra scaffolding.
• Active development with a roughly bi-weekly release cadence, though 2.0 introduced breaking API and session schema changes from 1.x.

Caveats

• Breaking changes in 2.0 affect the agent API, event model, and session schema; sessions are incompatible with ADK versions below 1.28.

Verdict

A solid candidate if you want explicit graph-based orchestration for Python agents without gluing everything together yourself. Probably not compelling enough to migrate if you’re already productive in a rival framework that covers similar ground.