Google ADK

Google Agent Development Kit (ADK) is an open-source framework for building, running, and deploying AI agents powered by large language models. Developed and maintained by Google, ADK provides a structured programming model for creating agents that can use tools, coordinate with other agents, and maintain state across complex workflows.

ADK supports multiple programming languages — Python, TypeScript, Go, and Java — making it accessible to a broad range of engineering teams rather than being limited to Python-first ML workflows. The framework is built around Gemini as the primary model but supports other providers including Claude (Anthropic), Vertex AI-hosted models, Ollama, vLLM, and any LiteLLM-compatible endpoint. This model flexibility distinguishes it from tightly coupled alternatives like LangChain's LangGraph (which is model-agnostic by design) or CrewAI (which defaults to OpenAI).

At its core, ADK organizes agents into three categories: LLM agents (which use a language model to reason and select tools), workflow agents (which execute deterministic patterns like sequential, parallel, or loop pipelines), and custom agents (fully user-defined logic). These can be composed into multi-agent systems where specialized sub-agents handle different tasks under an orchestrating parent agent — a pattern well-suited for complex pipelines such as research assistants, code review bots, or customer support automation.

The tool integration story is a notable strength. ADK supports function tools (plain Python/TypeScript functions), MCP (Model Context Protocol) tools, and OpenAPI-specified tools. This means agents can connect to REST APIs, local services, or the growing ecosystem of MCP-compatible servers without writing custom adapters.

For runtime, ADK ships with a built-in web interface for local testing, a CLI for scripted workflows, and an API server mode for integration with other services. Deployment targets include Google Cloud's Vertex AI Agent Engine, Cloud Run, and GKE — which makes ADK a natural fit for teams already operating in the Google Cloud ecosystem. That said, the framework itself is open-source and can run outside GCP.

Observability is handled through a logging subsystem and evaluation tooling that supports user simulation, allowing teams to test agent behavior against realistic scenarios before shipping. Sessions, state, and memory management are first-class concepts, with support for context caching, context compression, and session rewind — capabilities that matter when building long-running or resumable agents.

Compared to LangGraph, ADK provides a higher-level abstraction with more opinionated defaults and stronger Google Cloud integration. Compared to AutoGen or CrewAI, it offers broader language support and a more explicit runtime model. Teams deeply invested in the Google Cloud platform or building production agents with Gemini models will find ADK's deployment primitives and tooling particularly well-suited to their needs.