Reviewable Flutter Generation Workflows

Teams building Flutter clients for ADK agents can require the coding agent to produce four reviewable files before it writes UI code: AGENT_INTERFACE_NOTES.md, FRONTEND_USAGE_NOTES.md, FRONTEND_ARCHITECTURE_NOTES.md, and FRONTEND_DESIGN_NOTES.md. This gives the developer a concrete place to check the backend interface, expected user behavior, state model, and visual plan while the cost of revision is still low.

The useful adoption change is a gate in the coding-agent workflow. The agent reads the ADK codebase, drafts the notes, pauses for review, and only then generates the Flutter app. A cheap test is to run the workflow on two existing ADK agents with no frontend and check whether the notes correctly identify streaming behavior, APIs, tool calls, and required client state before code is generated.

A reusable Flutter frontend skill for ADK agents should include a post-generation checklist tied to failures that appear when the app runs on real targets. The checklist can cover macOS and iOS network permissions, markdown rendering with flutter_markdown, lint and formatting rules, sealed message types, chat auto-scroll, package:http networking for web builds, partial streaming event aggregation, and tool invocation display.

This is a concrete support layer for generated clients. The coding agent can run formatting and lint checks, scan for dart:io in web-bound networking code, verify entitlements, and include a small streaming-event test that assembles partial ADK events before adding completed messages to the session list. These checks target errors that a single code-generation pass can miss.

Developers using coding agents on unfamiliar ADK backends can keep a separate review loop for the reusable skill. One agent runs the current workflow and generates artifacts. A second conversation reviews the artifacts with the developer, records gaps, and updates the skill. After the update, the generated frontend is deleted and the workflow is run again.

This works best when the developer reads the generated notes and source files, runs the app, and directly fixes repeated failures. The next prompt asks Antigravity how to fold those fixes into the skill, so the next generated client starts with the newly learned rules. The workflow is useful for teams that need generated code they can explain and maintain after the agent session ends.