Google releases ADK 2.0: a graph-execution Python framework for agents

ADK 2.0 is a code-first Python framework requiring Python 3.11+, installed via pip install google-adk, with roughly biweekly releases. Its core addition is a Workflow graph execution engine supporting routing, fan-out/fan-in, loops, retries, state management, dynamic nodes, human-in-the-loop, and nested workflows. It also provides a Task API supporting multi-turn task mode, single-turn controlled output, and hybrid delegation, letting task agents serve directly as workflow nodes. Tooling supports built-in tools, custom functions, OpenAPI specs, and MCP tools, with a tool confirmation flow retained. The framework is optimized for Gemini but model-agnostic, and can be containerized on Cloud Run or scaled via Vertex AI Agent Engine. For Python developers building multi-agent apps, graph workflows make orchestration logic that once had to be hand-assembled a built-in capability, improving determinism and testability; cross-agent delegation scenarios can call the Task API directly instead of designing their own protocol. Teams on Google Cloud can reuse their Cloud Run and Vertex AI deployment paths, and projects already using MCP tools can reuse them without changes. For local development, adk run starts an interactive CLI, adk web launches a Web UI with evaluation and debugging, and adk eval runs evaluation sets. Teams using other languages can turn to companion Java, Kotlin, Go, and TypeScript versions, with community tools and third-party integrations housed in the adk-python-community repo.