Vibeleaderboard
← Back to Tools
Index / agent

Agent Development Kit (ADK)

https://github.com/google/adk-python
Visit google.github.io
Category
AI Agents
Rank
No. 154Tools index
Pricing
Open Source
Type
AGENT
Builder
google
GitHub
21.5k stars
Latest release
v1.39.1
Date

About

Google's open-source Python toolkit for building, evaluating, and deploying production-grade AI agents with full control over the agent loop.

What it does

ADK models an application as agents that hold instructions, tools, and behavior, then connects them through graph-based workflows. Its runtime supports routing, parallel branches, loops, retries, shared state, nested flows, human checkpoints, and structured delegation between agents. Developers can run projects from a command line or through a local web interface.

Why it's ranked here

The strongest case for ADK is breadth within one Python package. It combines direct agent construction with deterministic workflow graphs, multi-turn delegation, event-driven execution, evaluation extras, and local inspection tools. That scope is useful, but the major API, event, and session changes between generations make migration discipline essential.

What's good

Workflow graphs cover routing, fan-out and fan-in, loops, retries, dynamic nodes, state, nesting, and human approval. The task model supports both multi-turn delegation and controlled single-turn output. The package also separates many integrations into optional extras, while runnable samples demonstrate basic agents, authentication, artifacts, callbacks, long-running approvals, and cloud integrations.

Tradeoffs

The current generation breaks compatibility across the agent API, event model, and session schema. Its sessions work with only some earlier releases and fail with older ones. The base installation already carries a substantial dependency set, including web servers, telemetry, validation, authentication, graph rendering, and database support. Broader integrations add an even larger dependency surface.

How to use it well

Choose ADK for Python teams building agent systems that need explicit orchestration, delegation, retries, state, or human approval. Start with one agent, move stable control flow into workflow graphs, and install only the extras required by each deployment. Use the command line and local web interface during development. It does not replace infrastructure selection, model credentials, or external service setup.

Technical notes+

pyproject.toml uses flit_core.buildapi, requires Python 3.10+, declares the package as typed and OS independent, and defines separate extras for MCP, A2A, evaluation, databases, Google Cloud, Slack, OCI, and broader extensions. src/google/adk/__init__.py maps five public symbols to their defining modules and resolves those imports on attribute access through _lazy.accessors. contributing/samples/hitl/human_in_loop/main.py demonstrates asynchronous event streaming, persisted session identity, long-running tool correlation by call ID, resumed execution after approval, and OpenTelemetry export to Cloud Trace.

Observed

License
Apache License 2.0
Primary language
Python
Python support
Python 3.10 through 3.14 are declared; Python 3 only
Installation
Published as the google-adk package for pip; built with Flit
Interfaces
Python library, interactive CLI, and local web UI
Platform
Declared operating-system independent
Optional integration surfaces
Separate extras include MCP, A2A, evaluation, databases, Google Cloud, Slack, OCI, and extensions

Read from README.md, pyproject.toml, src/google/adk/__init__.py, contributing/samples/core/abort/__init__.py, contributing/samples/a2a/a2a_auth/__init__.py, contributing/samples/a2a/a2a_basic/__init__.py, contributing/samples/core/logprobs/__init__.py, contributing/samples/core/artifacts/__init__.py, contributing/samples/core/callbacks/__init__.py, contributing/samples/hitl/human_in_loop/main.py, contributing/samples/core/quickstart/__init__.py, contributing/samples/core/empty_agent/__init__.py, contributing/samples/core/hello_world/__init__.py, contributing/samples/integrations/gcs/__init__.py, contributing/samples/adk_team/adk_pr_agent/main.py.

What it can do

  • Build AI agents with custom logic

    Python code and agent specifications → Functional AI agent

  • Evaluate agent performance

    AI agent and test scenarios → Performance metrics and evaluation results

  • Deploy agents to production environments

    Trained AI agent → Production-ready deployed agent

  • Control agent decision-making loop

    Agent loop configuration and parameters → Customized agent behavior and execution flow

  • Debug agent execution

    Agent runtime data and logs → Debug information and error diagnostics

  • Monitor agent behavior in real-time

    Running AI agent → Live monitoring data and agent status

Tags

agentssdkgooglepythonllm

Tech Stack

Python

Media

Agent Development Kit (ADK)

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.