Blog

Vaadin and AI: three directions, starting with coding agents

By  
Miikka Andersson
Miikka Andersson
·
On Nov 16, 2026, 8:54:21 PM
·
In Product

Vaadin works on AI in three directions: tools that help coding agents build Vaadin applications, AI features for the applications you build, and a way for AI agents to use those applications. A series of three webinars on the first direction starts Wednesday, November 18, when Petter Holmström shows how to set up your coding agent for Vaadin.

Building with AI

If you let Claude Code or Codex write your Vaadin views, you may have seen it guess at APIs from older versions, rebuild a component Vaadin already has, or stop without knowing whether the change took effect. The Vaadin MCP server gives the agent the documentation for your Vaadin version. Agent skills are instruction files that teach it Vaadin's components and patterns. With both, the agent can use a ready-made component such as Grid instead of writing its own table.

After each change, the agent runs Dev Loop, a command-line tool that compiles the change, hot-swaps it into the running application or restarts it, and reports whether it loaded. Browserless UI tests check that the feature works. They are JUnit tests that run your views inside the JVM, with no browser or web server. Because a Vaadin view calls your Java services directly, a test that clicks Save also runs the form validation, the service and the database query behind it. A suite of browser tests can take minutes. Without a browser to start, browserless tests are fast enough for the agent to repeat the whole suite after every change. Rendering and browser behavior still need end-to-end tests.

AI features inside your applications

You can build AI features in Java with Spring AI, LangChain4j or your own LLM stack, and keep prompts and logic on the server. Vaadin's open-source AIOrchestrator class connects that backend to the chat components in your UI. It handles streaming responses, conversation history, file attachments and tool calling.

Vaadin 25.2 added AI controllers for Grid, Chart and forms. Your application's users can ask for a grid or chart in plain language. The LLM sees your database schema but never the rows, and it returns an SQL query and a configuration for the component. The query runs under the database account you give the controller, and that account should be read-only. The form controller fills in a form from a prompt, or from a document or image the user drops on it. Since Vaadin 25.3, the form marks each field the AI changed, and the user can check the value and revert it.

Applications agents can use

This direction is the least finished. Vaadin's Draiv project, now in alpha, lets an AI agent use a Vaadin application through the same UI a person uses. The business rules and permissions that the UI enforces apply to the agent too. Read Driving Vaadin apps with macros, automation, and AI for how it works, and sign up there for early access.

The first series: three webinars on coding agents

Set up your coding agent for Vaadin

Wednesday, November 18, 15:00 CET / 9 AM ET. Petter Holmström, Principal Software Engineer at Vaadin, sets up the MCP server, the skills and Dev Loop in a Vaadin 25 project in Claude Code or Codex, and explains why Vaadin built each one. At the end, the agent checks its work with browserless UI tests. Register.

How well do coding agents write Vaadin?

Wednesday, December 2, 15:00 CET / 9 AM ET. Vesa Nieminen, Principal Software Engineer at Vaadin, walks through VaadinBench, Vaadin's open benchmark for coding agents, and its October results. Each task gives the agent a job, such as building a view to a spec, and tests decide pass or fail. Vesa shows which models pass the most tasks, and the dollar cost of each try.

The session also compares Vaadin and Angular on one task, the vets page of the Petclinic demo app, with no MCP server or skills on either side. Each agent had 20 tries per framework. Codex passed 16 times in Vaadin and 5 times in Angular. Claude Code passed 17 times in Vaadin and 12 times in Angular (full results). Vesa also shows how to run VaadinBench yourself with a git clone and two commands. Register.

Can your agent run UI tests in milliseconds?

Tuesday, January 12, 2027, 15:00 CET / 9 AM ET. Simon Martinelli, Java Champion and Vaadin Champion, builds a feature with a coding agent, starting from a use case spec. The spec names the feature's business rules and the scenarios it must handle. The agent writes the Spring Boot and Vaadin code and checks each change with browserless tests. He also shows how to trace each test back to a business rule and find the rules that have no test yet. The session builds on his post on browserless testing for AI-generated applications. Register.

The other two directions in 2027

In 2027 we'll run webinars on AI features inside your applications and on applications agents can use. Until then, the AI page covers what you can build into your applications today.

We're already building in all three directions, and we want to know if we're building the right things. Where does Vaadin development slow you down, with an agent or without one? Where could we help you build with AI? What AI features do your users ask for? Tell us on the Vaadin forum, or join our webinars for a live conversation.

Start with Petter's session on Wednesday: save your spot.

Miikka Andersson
Miikka Andersson
Miikka is a versatile software engineering and product business professional with over two decades of experience. He joined Vaadin as a Technical Product Marketing Manager in mid-2024 to drive product growth and connect technical solutions with customer needs.
Other posts by Miikka Andersson