Google Launches an Agent SDK - Agent Development Kit

Google has launched an “Agent Development Kit” (Agent SDK) aimed at building deployable AI agents in the cloud, with built-in support for evaluation, tool integration, and multi-agent architectures. The push matters because agent frameworks have proliferated across the industry, but most still require extra work to make them production-ready—Google’s pitch is that deployment readiness is part of the foundation rather than an afterthought.

A key theme is operational readiness from day one. Instead of centering on local-only experimentation, the kit is designed to run remotely in cloud environments, and it includes mechanisms for evaluation alongside core agent capabilities. That emphasis on testing and deployment aligns with how teams typically adopt agent systems: they need repeatable runs, measurable performance, and a path to production rather than just demos.

The SDK also leans heavily into tool use and interoperability. From the start, it supports integrating tools from other ecosystems such as LangChain, and it includes “function calling” style built-in tools. Google Cloud’s existing tool ecosystem is positioned as a strength, with support for MCP tools and OpenAPI tools mentioned as part of the initial setup. There’s also an authentication system for tool access, and the framework is described as event-driven—an architectural choice that can make agent workflows more modular and responsive.

Another differentiator is multi-agent design. The kit is built around a multi-agent architecture from the outset, rather than treating multiple agents as an add-on. The GitHub documentation also points to core agent primitives such as state and memory, plus handling “artifacts,” suggesting the framework is meant to support more than simple chat-style interactions.

While the release is early—sample agents appear to return a GitHub 404 at the time of recording—Google has already published installation instructions and API references. The SDK is currently Python-only, which may disappoint developers looking for JavaScript support. Model support is also broader than a single vendor: the kit references Gemini models but also indicates compatibility with OpenAI models and Claude Sonnet through a “light LLM integration.” That matters because it reduces lock-in and lets teams compare agent behavior across model families.

The most consequential near-term question is how well the framework integrates with Google’s Gemini lineup, including Gemini 2.5 Pro. If Gemini 2.5 Pro training data incorporates agent-framework patterns, the result could be agents that work more effectively out of the box—an advantage similar to how teams benefit when model behavior is tuned to the surrounding tooling. The transcript frames this as an early-stage rollout, with room for strengths and weaknesses to emerge as developers build real agents.

For hands-on testing, the kit is available at github.com/google/adken python, and further coverage is expected as more videos and experiments roll out alongside other Google Cloud Next announcements, including new TPUs and an additional framework for agentic agent communication.