Agent Development Kit (ADK)

Agent Development Kit (ADK) is a code-first framework for building, evaluating, and deploying AI agents. Agent logic, tools, and orchestration are written directly in code, and the same model is available in Python, TypeScript, Go, Java, and Kotlin. It is optimized for Gemini but model-agnostic, deployment-agnostic, and compatible with other frameworks.

Version 2.0 is built around two classes. Agent defines an agent's instructions, tools, and behaviour; Workflow orchestrates agents and tasks as a graph of edges. The workflow runtime handles routing, fan-out and fan-in, loops, retries, state management, dynamic nodes, human-in-the-loop steps, and nested workflows, and a Task API adds structured agent-to-agent delegation with multi-turn tasks, single-turn controlled output, and task agents used as workflow nodes. The 2.0 release changed the agent API, event model, and session schema: sessions it writes are readable by 1.28 and later, but not by earlier 1.x versions.

Locally, adk run opens an interactive session against an agent directory and adk web serves a development interface for testing, debugging, and demonstrating one or many agents. adk eval runs an agent against a stored eval set. Agents can also be described declaratively through Agent Config instead of code. Python 3.10 or later is required, and companion constraints files pin transitive dependencies per Python version.

Features

• Workflow runtime: a graph-based execution engine with routing, fan-out/fan-in, loops, retry, state management, dynamic nodes, human-in-the-loop, and nested workflows
• Task API: structured agent-to-agent delegation with multi-turn and single-turn modes and mixed delegation patterns
• Multi-agent composition: specialized agents combined into hierarchies for larger applications
• Tool ecosystem: pre-built tools, custom functions, OpenAPI specs, MCP tools, and existing tools brought in as-is
• Tool confirmation: a human-in-the-loop flow that guards tool execution with explicit confirmation and custom input
• Agent Config: agents defined without writing code
• Evaluation: adk eval scores an agent against a stored eval set
• Development interface: adk web for testing, debugging, and showing agents locally
• Deployment: containerized deployment on Cloud Run or Vertex AI Agent Engine
• Multiple languages: Python, TypeScript, Go, Java, and Kotlin implementations of the same framework
• Machine-readable context: llms.txt and llms-full.txt provide summarized and complete project context for a coding agent