Google ADK

Google ADK is an AI agent framework for developers building applications that carry out multi-step tasks. Agents can run on your own infrastructure and use locally running models through adapters, or connect to hosted providers such as Gemini.

• Workflow control: Define execution paths with parallel tasks, loops, retries and pauses for human input, combining explicit logic with model reasoning.
• Agent delegation: Split work among specialized agents that can handle ongoing exchanges or single-turn tasks with controlled output.
• Tools and approval: Connect custom functions, MCP tools and OpenAPI services, and require confirmation before selected tools execute.
• Testing and evaluation: Use the built-in development UI to interact with agents, debug their behavior and evaluate results before deployment.
• Conversation context: Manage sessions, memory, tool results and artifacts with filtering of irrelevant events, summaries of older turns and token tracking.

ADK is available in Python, TypeScript, Go, Java and Kotlin; the Python repository uses Apache 2.0. It supports local Docker deployment and self-hosted infrastructure, alongside Google Cloud deployment through Agent Runtime, Cloud Run or GKE. It's optimized for Gemini but supports other models and providers. Hosting an agent yourself doesn't keep its model requests local if you connect it to a cloud model service.