google/adk-python
Google's Python toolkit lets you code AI agents and wire them into multi-agent workflows
ADK (Agent Development Kit) is an open-source Python framework from Google for building, testing, and deploying AI agents entirely in code. It's optimized for Gemini models but works with other models and frameworks too. Version 2.0 adds a graph-based workflow engine and a Task API for agents to delegate work to each other.
- agent
- agentic
- agentic-ai
- agents
- agents-sdk
- ai
- ai-agents
- aiagentframework
- genai
- genai-chatbot
- llm
- llms
What it does
- A workflow runtime lets you connect multiple agents in a graph, handling routing, fan-out/fan-in branching, loops, retries, and human-in-the-loop checkpoints
- The Task API enables structured agent-to-agent delegation, and agents can be equipped with prebuilt tools, custom functions, OpenAPI specs, or MCP tools
- Agent Config lets you build agents without writing code, and Tool Confirmation adds a human approval step before a tool runs
- Install with pip install google-adk (Python 3.10+ required), then run locally with adk run or test in a browser with adk web
- Agents can be containerized and deployed on Cloud Run or scaled via Vertex AI Agent Engine
- Version 2.0 introduces breaking changes to the agent API, event model, and session schema, making it incompatible with older 1.x session data
Why it matters
It gives developers a ready-made structure for building AI agents instead of writing orchestration logic from scratch, cutting development time. It also makes complex, multi-agent systems manageable through code and visual workflow graphs, which matters for anyone shipping agentic apps to production.
Terms in this repo
- Agent · an AI unit defined by instructions and tools that can act on its own
- Workflow · a graph structure that connects agents or tasks to control execution order
- Task API · a mechanism for agents to hand off and receive work from each other
- MCP · a standard protocol for connecting AI to external tools and data
- human-in-the-loop · a step where a person reviews or approves an automated action
Repository description (English)
An open-source, code-first Python toolkit for building, evaluating, and deploying sophisticated AI agents with flexibility and control.
Open on GitHubProject siteCoverage
Trending repos
- vorssaint/vorssaint-utilsOne free menu bar app replaces a dozen paid Mac utilities
- MadsLorentzen/ai-job-searchAn open-source framework that runs your entire job search through Claude Code, built by someone who actually got hired using it
- rohitg00/ai-engineering-from-scratchA free hands-on curriculum that teaches AI by having you build every piece from scratch, from math to shippable agents
- AgriciDaniel/claude-obsidianAn open-source second brain that lets Claude Code file your notes into Obsidian while you keep full ownership of the files
- openclaw/openclawA personal AI assistant that lives on your own devices and talks to you through the messaging apps you already use
- tashfeenahmed/freellmapiAn open-source router that stitches 34 free LLM tiers into one API key so your app never hits a rate limit
- Alishahryar1/free-claude-codeA local proxy that lets coding AI agents run on 49 free or cheap model providers instead of one paid service
- freestylefly/awesome-gpt-image-2A library of 532 reverse-engineered prompts that turn GPT-Image2 into a predictable image-making tool
Latest from METAL LAB
- Alibaba's Wan3.0 reveals pricing: $6 for 30-second 1080p video
- Thomson Reuters built its own AI instead of renting one, for $40 million
- Google's ADK adds automated evaluation for live voice agents
- ChatGPT Work now turns spoken thoughts into finished documents
- 100-Rally Match Against World No. 15... The Robot on Court Was Unveiled Three Days Earlier