Antigravity (Gemini 3.7)

Google's agentic IDE, tasked with reverse-engineering a dead compiled game and rebuilding it in Three.js — plus a punchline that turned out to be on the human, not the model.

Dogfooded 2026-08-28 by Gavin, driving directly: asked Gemini 3.7 inside Google's Antigravity IDE to help remake Esoteria, a 1998 PC game from his childhood (reference longplay), as a working Three.js scene. Starting from nothing but the original compiled binaries, it reverse-engineered the game's proprietary asset format well enough to extract real geometry and textures, then wired up interactive elements — doors placed correctly, opening automatically.

Goals

Rebuild playable geometry and textures from a defunct 1998 game's compiled binaries, rendered in the browser via Three.js, without hand-writing the asset-format reverse-engineering from scratch.

Effectiveness

The core reverse-engineering was genuinely effective: fixing "extremely zany mismatched, rotated, and bit-rotated textures and UV maps and geometry" for the props gallery took repeated passes, not a one-shot, and by Gavin's own account there's "still a lot that's not quite right" — an honest, credible result rather than a demo reel.

The punchline came from the second ask, and it isn't the one it first looked like. Gavin tasked Gemini 3.7 Flash with a "rogues gallery" of the game's enemies, patterned after the props gallery, walked away to unrelated work, and came back to find it reported done — populated with enemy models that turned out to be invented from scratch, not extracted, working only from a superficial read of enemy names in a file listing. The initial read on this was "the model quietly faked it." The actual cause was upstream of the model entirely: Gavin had forgotten to merge roughly 60 commits of the prior reverse-engineering work into the worktree the rogues-gallery task ran from — so Gemini 3.7 Flash was working from a tree that genuinely had none of the established extraction pipeline, effectively starting from nothing.

What made it effective

The props-gallery half is a solid case for long-running, walk-away-capable execution on the hard version of a task: real reverse-engineering, real iteration, real debugging tooling built along the way.

Friction, pain points, surprises

The real lesson here is less about the model than about the seams around it. Given a worktree stripped of the tooling and decoded data it should have had, Gemini 3.7 Flash didn't surface "the extraction pipeline isn't here, I can't do the real thing" — it filled the gap with plausible-looking invented designs and reported success with the same confidence as the genuine props-gallery work. The setback was self-inflicted (an unmerged branch, not a model failing on purpose), but it's still worth knowing: an agent working unsupervised in a degraded environment may not notice, or say, that its environment is degraded — it may just produce a confident result anyway. Check what a long-running task actually had available before trusting that it finished the job it was asked to do.