Skip to content

Game-Agent Open StandardBuild games that become verifiable benchmarks.

One deterministic game contract for human play, agent evaluation, and independent verification of exact runs.

Game-as-a-Benchmarkv1.0 stable product-neutral SDKTypeScript + Python

Why Game-as-a-Benchmark ​

As agents and world models move from answering questions to acting over time, evaluation needs environments where decisions change what happens next. Games make planning, adaptation, memory, cooperation, and failure visible inside a controlled world.

A Game-as-a-Benchmark product is both playable and measurable. People and agents reach the same versioned rules; scores summarize performance; portable evidence preserves how each result was produced.

Read the thesis →

One SDK, two builders ​

For game developers

Build for people. Stay agent-ready.

Use one product reducer for the renderer, authoritative host, bots, tournaments, and replay. Compose reusable mechanisms without surrendering your rules, content, presentation, or commercial policy.

Build your first game →
For benchmark builders

Turn play into defensible evidence.

Expose structured observations and legal actions, evaluate single or multiple agents, and publish the result with evidence of the exact run. You retain task design, score meaning, held-out content, and claims.

Build an evaluation →

The verifiable arm ​

Reproducing an evaluation repeats its inference cost and produces a new stochastic sample. It does not check the published run.

GAOS preserves canonical reducer inputs, results, signatures, and chains so an independent verifier can re-simulate the exact recorded run:

text
human or agent play
        ↓
signed portable evidence
        ↓
historical reducer + semantic adapter
        ↓
independent verification

GAOS makes the historical verifier portable as a product-owned, content-addressed kit. GAOS checks and executes the kit; the product preserves and publishes it; an independent manifest, catalog, or allowlist authorizes its digest. If required verifier material is unavailable, the result is unverifiable—not silently trusted. v1.0 freezes that evidence contract while keeping product-specific hosted adapters outside the standard SDK.

Understand the verification boundary →

Built with GAOS ​

Zonoid

The first production game built with the GAOS SDK.

A strategy game for humans and AI agents, built around prediction, planning, and judgment.

Visit Zonoid →

See all games built with GAOS →

Play the demos → · See the architecture → · Browse capabilities →

Game-as-a-Benchmark: one product reducer for human play, agent evaluation, and verifiable exact runs.