Antigravity’s strongest weekly signal is reusable coding skill work, not benchmark claims

The strongest case study turns a one-off coding session into a reusable Antigravity skill. The author needed a Flutter client for a Python agent built with Google Agent Development Kit, despite little Python experience and no prior ADK work. The useful method was staged output: interface notes, usage notes, architecture notes, design notes, then code generation and testing.

The workflow used two agents. A coder agent ran the skill and produced artifacts. An author agent helped inspect gaps and update the skill. After each run, the author deleted the generated files and reran the process. This made the skill a record of fixes, decisions, and domain learning.

The concrete result was 13 iterations before a shareable flutter_frontend_for_adk skill. The reported fixes were ordinary engineering problems: gitignore behavior, review pauses, macOS and iOS network permissions, markdown rendering, linting, sealed message types, chat scrolling, web crashes caused by dart:io, partial event aggregation, and tool invocation display.

A second item makes the same practical point through a shipped game example. DashLander was built with Flutter, Flame, Antigravity, and Gemini tools. The argument is that one Dart codebase gives the agent compiler and analysis-server feedback while keeping web, mobile, and desktop builds close together.

The claim is grounded in an end-to-end build, not a controlled evaluation. The authors report an initial Flutter and Flame game in about five minutes, followed by about 100 more prompts. Much of that later work went into code organization and tests. The game also needed human inspection tools, including debug overlays and replay checkpoints, because physics, collision logic, and replay timing were hard to trust by reading generated code alone.

The clearest engineering lesson is about scope control. Challenge Mode used ghost replays of past high scores, which avoided real-time multiplayer infrastructure. A replay bug caused by millisecond-level scheduling drift was fixed by storing complete lander state at thruster events.