I gave Gemini 4 Argon in Antigravity and Claude Code the same messy side project, and only one of them left it better than it found it

An evaluation tasked Google's Gemini 4 Argon and Claude Code with improving a smart home dashboard project for a jailbroken 2015 Amazon Kindle Paperwhite. This side project included a Node 20 websocket proxy and a KUAL extension for the Kindle, with specific constraints like requiring ES5 JavaScript for the Kindle's ancient browser and Linux line endings for its launch script. Both AI agents successfully maintained these legacy requirements and fixed a proxy issue, albeit using different technical solutions.

A key distinction emerged in their testing methodologies. Claude Code generated 63 tests, notably including one for the hixie-76 WebSocket draft essential for the Kindle's old browser. Gemini 4 Argon, by contrast, produced only four tests. This difference was evident during real-world testing: when connected to a Home Assistant instance missing calendar data, Claude Code's proxy logged each missing calendar separately and continued processing, demonstrating specific error handling and resilience.

Conversely, Argon's proxy logged a single, generic error for the entire list of missing items. This suggests that while both agents could address immediate functional problems, Claude Code's more extensive testing regimen and granular error reporting point to a more robust approach to code improvement. This method appears better suited for anticipating and managing diverse failure conditions within a complex, legacy system, going beyond basic functionality fixes.