Inspiration

Most AI benchmarks test one answer. We wanted to test sustained behavior: planning, adapting, recovering from mistakes, and handling consequences. Games make that measurable through shared rules, repeatable scenarios, scores, and complete decision traces.

Simultaneous turn-based grids give humans and agents the same state and action window, removing reaction speed, network latency, and request order from the contest. Each turn instead tests prediction, spatial reasoning, coordination, and opponent modeling. GAOS grew from this idea while building Zonoid Labs AGI Arena.

What it does

GAOS is an open-source TypeScript and Python toolkit for deterministic, simultaneous turn-based grid games. One authoritative engine provides movement, recursive settlement, pathfinding, FOV, pushing, projectiles, transport, gates, triggers, pickups, scoring, solving, and replay verification. Agent environments, model drivers, MCP tools, and extensible CLIs add legal actions, seeded episodes, transcripts, and batch evaluation without graphical UI automation.

Zonoid was central to production: as the game evolved, GAOS generalized, implemented, and tested the reusable capabilities needed to meet its new requirements.

How we built it

We extracted Zonoid's product-neutral rules into an Apache-2.0 SDK and built one deterministic reducer for gameplay, solvers, replays, and agents. We added TypeScript and Python interfaces, provider-independent model drivers, MCP and CLI integrations, tests, packages, releases, CI, and a VitePress site.

Codex with GPT-5.6 Sol supported the complete production loop detailed below. People retained final judgment over scope, gameplay, visual approvals, and releases.

Challenges we ran into

The first challenge was the SDK boundary: reusable mechanics, settlement, scoring, solving, and agent contracts belong in GAOS; Zonoid-specific characters, campaigns, levels, and seasonal rules do not. The second was recursive resolution. A move can trigger collisions, pickups, switches, gates, transport, or more movement, so deterministic settlement must continue in ordered waves until the state is quiet. We also had to keep model and CLI extensions provider-neutral.

Accomplishments that we're proud of

We shipped a public Apache-2.0 repository, a documented v0.9.1 release, prebuilt npm and Python packages, detailed mechanism pages, replayable transcripts, solver and scoring support, MCP tools, and extensible model and CLI drivers. Its deterministic engine is backed by 98 TypeScript and 13 Python tests. Most importantly, humans, agents, solvers, replays, and Zonoid share one authoritative reducer—agent support is part of the game contract, not a later automation layer.

What we learned

AI-native games need a clear state-and-action contract, not screen automation. Determinism also requires explicit ordering, IDs, seeds, versions, and recursive effects—not simply removing randomness. Separating the SDK lets Zonoid focus on authored experiences while GAOS becomes easier to test and extend. The two projects improve each other through a continuous requirements-and-validation loop.

What's next for Zonoid Labs AGI Arena

We will open registration and downloads, expand held-out tasks, multiplayer evaluation, and leaderboards, and report richer measures such as consistency, efficiency, recovery, and failure modes. As Zonoid introduces new gameplay and research needs, reusable solutions will continue to strengthen GAOS. We will also move the SDK toward 1.0, welcome community contributions, and publish research data from Zonoid while keeping its product content outside the SDK.

Required URLs and identifiers

Public code repository: https://github.com/yugao-gaos/GAOS-TurnBasedGrid-SDK Documentation: https://yugao-gaos.github.io/GAOS-TurnBasedGrid-SDK/

Live game and prebuilt download: https://zonoid.ai Register a free account.

Latest SDK release: https://github.com/yugao-gaos/GAOS-TurnBasedGrid-SDK/releases/tag/v0.9.1 Codex /feedback Session ID: 019f8458-7a8d-7010-9227-500c99df5e04 License: Apache License 2.0

Built With

  • codex
  • elevenlabs
  • godot
  • seedance
  • tripo3d
  • world-labs-marble
Share this project:

Updates