Gemini 4: Argon, Early Access Review

Google handed out early access to Gemini 4, codenamed Argon, through its Develop Experts program, and the tester threw three very different jobs at it: a mobile game, a RocketFlag UI refresh, and a deep backend audit. The verdict is mixed, but not vague. Argon looked strong when the task was about planning, generating code across multiple files, and getting something working fast. It looked weaker when the work depended on seeing what a human would see, or when concurrency and state got messy.

The game project is the easiest place to see the model’s strengths. The build started as a rail shooter inspired by StarFox, with level progression, increasing difficulty, barrel rolls, bombs, gates, and charge shots with lock-on. Argon helped shape the app, fixed bugs in one shot when they were pointed out, and even used debug logs and crashes pulled from a phone to repair issues automatically. The result was described as genuinely fun, which is not something every mobile-game prototype gets to claim.

RocketFlag was where the rough edges showed up. The first design pass was noisy and cluttered, and the review says Argon leaned too hard into generic SaaS aesthetics before several rounds of direction cleaned it up. Still, it moved fast: it prototyped HTML-only pages for quick feedback, audited the frontend structure, and helped turn up a stack of problems, from a 3-column layout drift to a 2.3-second blocking animation on every flag toggle. With a Playwright loop in place, 14 targeted refinements landed in single passes, and 91 of 98 E2E scenarios stayed green on the first CI run.

The same pattern showed up in the audit work. A larger backend review produced 32 ranked findings across 5 waves of production hardening after nearly 1,300 agent steps. Some issues were fixed cleanly on the first try, including cache key collisions, cross-project IDOR, missing HTTP limits, Stripe webhook race conditions, and invite seat-cap races. But the useful bit came from forcing the model to review its own work: that caught production-breaking flaws like an infinite Pub/Sub reconnect loop, cross-node cache staleness, invite privilege escalation, and loose route matching.

The article’s real argument is that Argon works best when the harness keeps it honest. Static checks and test suites helped a lot, but live logs, browser verification, and adversarial self-review were what exposed the blind spots. That’s the story here: Google has a serious model, but the wrapper around it may matter almost as much as the model itself.