An AI agent building framework developed and published by Google. Released in April 2025, it provides SDKs for Python, TypeScript, Go, Java, and Kotlin. Rather than targeting one-off chatbots, ADK focuses on assembling “multi-agent” setups where several agents divide the work among themselves. It covers external tool connectivity (MCP), an agent-to-agent communication standard (the A2A protocol), streaming that includes audio and video, and an evaluation framework for measuring the quality of the agents you build. It is open source under the Apache 2.0 license, so the framework itself is free to use. A major characteristic is that the agents you build can be deployed from local containers all the way to Google Cloud’s Agent Runtime, Cloud Run, and GKE using the same code.
Key Features
- Designed around multi-agent composition: Define specialist agents by role and combine them hierarchically. A graph-based workflow execution engine lets you mix deterministic control ─ branching, loops, retries, nesting ─ with AI-driven decision making
- Rich tool integration: In addition to built-in tools, custom functions, and automatic generation from OpenAPI specifications, it supports connecting external tools over MCP (Model Context Protocol), so existing APIs and internal systems can act as the agent’s hands and feet
- Agent-to-agent communication via the A2A protocol: Agents built separately can talk to each other through a shared standard, which makes it easier to connect agents built by different teams within an organization
- Streaming and live conversation: Supports bidirectional exchanges that include audio and video, not just text, so it can be used where real-time responsiveness matters
- Evaluation and debugging facilities: You can evaluate not only the agent’s final response but the execution trace ─ which tools were called and how. A web-based development UI, a visual builder, and logging, metrics, and tracing are also provided
- Model choice: Optimized for Gemini, but the framework itself is model-agnostic. Models from Claude, OpenAI, Ollama, and others, as well as your own models, can be combined
Pricing
| Category | Cost | Details |
|---|---|---|
| ADK itself | $0 (Apache 2.0 license) | All SDKs, the development UI, and evaluation features are free to use |
| Model usage | Usage-based, billed by each model provider | API charges for Gemini API / Vertex AI or third-party models are separate |
| Cloud runtime | Usage-based, billed per service | Execution costs for Agent Runtime, Cloud Run, GKE, and similar are separate |
Pricing is as of August 2026. The framework itself is free, but real-world operation incurs model API and infrastructure costs. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- The framework itself is open source and free, so you can try it without worrying about license fees
- Broad language coverage ─ Python, TypeScript, Go, Java, Kotlin ─ makes it easy to match an existing team’s stack
- You can move from local development to production deployment on Google Cloud with the same code
- The evaluation framework and tracing come as standard, so you can measure quality instead of settling for “it seems to work”
- Because it builds on the industry-wide MCP and A2A standards, it connects easily with third-party tools and other teams’ agents
⚠️ Cons
- It is a code-first framework and not suited to people who want to build agents without writing code (there is a mechanism for defining agents through configuration files, but serious use requires programming knowledge)
- The breadth of features means many concepts to learn, so the initial ramp-up takes time
- Integration with Google Cloud is the most thorough, so running on other clouds means building more of the glue yourself
- Development moves quickly and structures can change between versions (the Go edition introduced significant changes in v2.0)
- Even though the framework is free, model API and infrastructure costs accumulate in production
Comparison with Similar Services
| Item | ADK | LangGraph | CrewAI | OpenAI Agents SDK |
|---|---|---|---|---|
| Provider | LangChain | CrewAI | OpenAI | |
| Main languages | Python / TypeScript / Go / Java / Kotlin | Python / JavaScript | Python | Python / TypeScript |
| License | Apache 2.0 (open source) | Open source | Open source | Open source |
| Strength | Multi-agent plus end-to-end cloud operation | Fine-grained state control via graphs | Easy role-based team composition | Integration with OpenAI models |
| Recommended models | Gemini (others supported) | Not tied to a specific model | Not tied to a specific model | Primarily OpenAI models |
Who Is It For
- Developers who want to build systems where multiple agents share the work, rather than a single chatbot
- Teams already using Google Cloud or Gemini who want to build an agent platform as a natural extension
- Environments where agents must be implemented in something other than Python (Go, Java, TypeScript)
- Engineers who want to continuously measure agent quality on the way to production
- Organizations that want to secure future interoperability by following standards such as MCP and A2A
Summary
Agent Development Kit (ADK) is a Google framework designed so that AI agents are not merely built and forgotten, but carried into operation with evaluation, debugging, and deployment included. It covers the elements today’s agent development calls for ─ multi-agent structures, MCP, A2A, streaming ─ and supports a wide range of languages. On the other hand, it is a code-first tool and the learning cost is not small. A realistic approach is to run a single agent using the official quickstart first, then expand step by step into tool integration and multi-agent composition.