Google adds open source framework for building agents to Vertex AI

Google is adding a new open source framework for building agents to its AI and machine learning platform Vertex AI, along with other updates to help deploy and maintain these agents.

It unveiled the open source Agent Development Kit (ADK) at its annual Google Cloud Next conference, saying it will make it possible to build an AI agent in under 100 lines of Python code. It expects to add support for more languages later this year.

[ Related: Google Cloud Next ’25: News and insights ]

ADK is built on the same framework that powers Google Agentspace and Google Customer Engagement Suite (CES) agents.

VP of Cloud AI Saurabh Tiwary, said developers will be able to use it through the entire agent development lifecycle, starting with shaping how agents think, reason, and operate within guardrails, and to select which LLMs they are running in the backend. Interaction with agents won’t be limited to text: ADK also provides bidirectional audio and video capabilities, he said.

Agent Development Kit to help build multi-agent systems faster

Gartner vice president analyst Jim Hare said he expects the ADK will help enterprises accelerate the build-out of in-demand multi-agent systems.

Making the framework open source will free up Google’s engineering resources, Hare said, as it will enable the company to leverage a community of developers to enhance and maintain the code base. “At the same time, Google gains by showing the developer audience it wants to help enterprises benefit from using agentic AI using open software and tooling,” he said.

Duncan Van Kouteren, an analyst at Nucleus Research, said that the open-source framework will also create a natural path for enterprises to move to Google’s paid cloud services. “Google anticipates that once developers build with their framework, they’re more likely to use Google Cloud for deployment,” he said.

Some of Google’s cloud rivals have similar ambitions, including IBM with its open-source BEE Agent Framework and Microsoft with AutoGen.

The ADK also supports the Model Context Protocol (MCP). Anthropic created the protocol to enable agents to interact with data sources and LLMs, and it’s rapidly gaining support from other software vendors.

Other Vertex AI updates

Google has also added a collection of ready-to-use, pre-built agent patterns and components named Agent Garden to accelerate the model development process.

It’s a strategy already followed by Salesforce and Microsoft, which introduced ready-to-use agent templates and components, in their Agentforce and Copilot offerings respectively.

Other updates to Vertex AI include a fully managed runtime engine to help developers deploy agents in production and maintain control over their behavior. Agent Engine is intended to handle infrastructure management and help with tasks such as rebuilding an agent or an agent system when moving it to production. Developers will be able to use Agent Engine to deploy agents or systems built on any framework, including ADK, LangGraph, and Crew.ai among others, on Google Cloud, Tiwary said.

Agent Engine can also help agents retain the context of sessions as it supports short-term memory and long-term memory, he added. This means that applications integrated with such agents will be able to recall past conversations and user preferences. Further, he said, it can be used with Vertex AI’s evaluation tools to improve agents’ functionality and performance.

Another option available to developers is to connect to Agentspace via AgentEngine. This is a way for enterprises to pass on agents to employees while maintaining control over them, Tiwary said.

Van Kouteren sees Agent Engine as a way to eliminate the overhead of managing the infrastructure to run agents in production.

“It handles all the behind-the-scenes work such as security, scaling, and monitoring, so enterprise teams can focus on what their AI agents do rather than worrying about how to keep them running properly,” he said.

Google said it intends to expand Agent Engine with computer-use and code-execution capabilities, and a dedicated simulation environment so developers can rigorously test agents with diverse user personas to ensure reliability in production.