Google open-sourced a Python framework for building agents

The conversation

On 9 April 2025, at Google Cloud NEXT, Google introduced the Agent Development Kit (ADK), describing it as "an open-source framework from Google designed to simplify the full stack end-to-end development of agents and multi-agent systems." The announcement states the framework is "Pythonic" in design, gives developers "precise control over agent behavior and orchestration," and already powers agents inside Google's own Agentspace and Customer Engagement Suite products. It shipped the same day as the separate Agent2Agent protocol, already on file, though the ADK announcement text does not itself describe that pairing.

What the documents show

Google's own repository for the Python implementation is released under the Apache License, Version 2.0, a permissive license letting any organization use, modify and redistribute the framework's code without a separate agreement with Google. The announcement states ADK supports "multiple LLMs (including Gemini, Anthropic Claude, and others via LiteLLM)," meaning the framework's stated design does not require Google's own models, a claim about compatibility rather than an independent test of how each supported model performs inside it. As retrieved 16 September 2026, the project's documentation site still describes ADK's purpose in close to the same words as the launch announcement, a toolkit to "build powerful multi-agent systems."

The system boundary

ADK's own materials describe it as a development framework, not a hosted runtime or a communication protocol: it is code a developer writes and runs, while A2A, launched the same day, is the wire format two already-running agents use to talk to each other, and MCP, already on file, is the format an agent uses to reach a tool or data source. The announcement lists interaction surfaces the framework offers a developer directly, a CLI, a Web UI, an API Server and a Python API, none of which is itself a network protocol that other agents outside the developer's own code would need to implement.

Where it fails

Because the April 2025 announcement does not name a version number or claim a measured performance result, a reader cannot cite it for a benchmark figure; it describes design goals and available interfaces as Google's own claims about what the framework supports. The announcement's claim about Claude and other non-Google models depends on LiteLLM, a separate third-party compatibility layer the announcement names but does not itself document, so a specific outside model's actual behavior is a fact to verify in that layer's own documentation rather than assume from Google's summary.

• Does a claimed multi-model or multi-framework compatibility rest on the vendor's own code, or on a named third-party layer?
• Is a given multi-agent system built with ADK, calling A2A, or both, and does the source say which?
• What license governs a framework's code, and does that license extend to models it merely calls out to?

ADK and A2A shipped on the same day from the same company but solve different problems, one a library a single developer runs, the other a protocol between separate agents, and Google's own materials keep that boundary even where a summary might collapse it.