Here’s the ADK documentation →

ADK is an open-source agent development framework that enables the creation, debugging, and deployment of reliable AI agents at an enterprise level. Compatible with Python, TypeScript, Go, Java, and Kotlin, developers can go from concept to coded ADK agent in mere minutes. Utilizing preferred AI-enabled developer environments allows for tasks such as scaffolding, building, testing, evaluating, and deploying agents using the Agents CLI.

A key feature of ADK is its ability to manage deterministic code alongside adaptive AI reasoning, enabling the orchestration of intricate tasks through structured graph-based architectures. These architectures provide explicit execution paths and predictable outcomes, ensuring efficient and reliable execution of complex tasks.

With the release of ADK 2.0, new capabilities such as prompt-based agent creation and tool call integration have been introduced. These enhancements allow for the progression to multi-agent orchestration, graph-based workflows, performance evaluation, and deployment on top-tier enterprise services for scalability, reliability, and throughput.

ADK's open integration partners facilitate connections between agents and existing applications, a variety of AI models, and additional functionalities like data access, resilience enhancement, and performance evaluation. The framework's commitment to inclusivity is evident in its design, which allows both human developers and AI to contribute to agent creation. Developers can even connect their favorite coding assistants to ADK's developer skills and AI-aware resources to generate agents swiftly.

ADK supports a wide range of generative AI models, providing easy access to Gemini as well as other leading models. Adapters are available to connect with numerous models and providers, including locally operated models. For enterprises, ADK enables connections to hosted models on services like Google Cloud, which offers a diverse range of models and allows for close management of performance, reliability, security, access, safety, and costs.

The framework prioritizes building professional, production-grade agents without necessitating extensive code. Its foundational agent structure is designed for ease of creation and flexibility, allowing for extensions and complexity additions as needed. ADK offers a range of development tools for interacting with built agents and AI-powered tools for creating ADK agents.

Context management is another area where ADK excels. The framework treats context like source code, assembling sessions, memory, tool outputs, and artifacts into a structured view where every token has purpose. ADK automatically filters irrelevant events, summarizes past conversational turns, lazy-loads artifacts, and tracks token usage. This efficient context management by default, while still providing controls for customizing context management based on specific task requirements.

With a focus on deploy-anywhere flexibility, ADK can be containerized and run on your own infrastructure or deployed with a single command to Google Cloud. When deployed via Agent Runtime (Agent Platform), Cloud Run, or GKE, agents inherit managed infrastructure, built-in authentication, Cloud Trace observability, and enterprise-grade security, all without requiring modifications to the agent code. This enables local development and global scaling.