Flutter Akure 2026: Your Coding Agent Is Flying Blind — Giving AI agents eyes and hands with Marionette

AI can already write Flutter. Screens, navigation, forms, buttons, API calls: a coding agent will happily generate all of it. The talk opens with the question that comes next: does it actually work? Today the agent can read source code, dependencies, tests and error output, but not the running app. So it guesses, and a human closes the feedback loop by looking at a screen and saying "this is wrong".

The second part introduces Marionette, an MCP server that exposes a running Flutter app as tools. It sits between any MCP client (Claude, Cursor, Gemini) and your app, talking to it through the Dart VM Service. The session walks through what happens on each tool call, the small toolset (connect, getinteractiveelements, tap, enter_text, scroll_to, take_screenshots, get_logs, hot_reload), the three-step setup, and how Marionette complements the official Dart & Flutter MCP server: one helps build the app, the other lets the agent use it.

The heart of the talk is four live demos, each a question for the agent. Can you see my app? Can you use my app, logging in and navigating to the dashboard? Can you find what's broken, reproducing a planted bug, diagnosing it, fixing it and hot reloading? Can you verify your own work, implementing form validation and testing every state?

It closes on what this means: the shift from a coding assistant that generates, modifies and explains to a development agent that also runs, observes, interacts, debugs and verifies. Marionette is a new layer, not a replacement for unit, widget and integration tests. Keep your tests, and give your agent eyes, hands and feedback inside the app it built.