Proof, not promises

Engine benchmarks

We validate Hexaliz Engine by building complete games with it in isolated environments — real projects, start to finish, by AI agents with human direction. 5–10 benchmark titles are planned; results are published here honestly, including what didn't work.

1

One prompt

A fixed game brief with defined scope.

2

Standalone run

A plain AI agent, no engine — the baseline.

3

Hexaliz run

Same agent, inside Hexaliz with MCP + skills.

4

Compare

Assets, logic, completeness, time.

The thesis: the engine's project-aware AI harness — chained skills, MCP tools, readable game state — produces better assets, more coherent logic, and more complete games than the same model working alone.

#1Complete2025-07

SubnauticaLite

Underwater survival

SubnauticaLite gameplay

The prompt

“Build a minimal underwater survival game: swim, manage oxygen, collect resources, avoid a patrolling creature, reach the surface to win.”

What it exercised

· Full game loop authored by AI agents + human direction
· Python ECS scripting (@system) with hot-reload
· Asset generation + import via Hexaliz MCP tools
· Open formats end-to-end (TOML scene, PNG/WAV assets)
· Browser export — same assets, HTML/three.js build

Results

✓ Minimum playable loop verified: swim + oxygen + resources + creature AI + win/lose states
✓ Runs both native in Hexaliz Engine and as a browser export from the same open assets
✓ Findings fed straight back into the engine (MCP + onboarding improvements)

Standalone vs Hexaliz

Side-by-side comparison captures coming — the baseline run is being documented.

Play in browser — soonHexaliz Engine · Python ECS · TOML/JSON assets · three.js export

More benchmarks are in the pipeline — different genres, same methodology.

Follow along by creating a free account — we announce each published benchmark.