Agentic play

GAOS turns a product-owned reducer into a structured environment for local policies, model drivers, and MCP-capable agent CLIs. Agents reach the same rules and canonical actions as human players.

Environment contract

AgentEnvironment advances one deterministic tick at a time:

import { AgentEnvironment } from '@yugao-gaos/gaos-sdk/engine';

const env = new AgentEnvironment({
  reducer,
  level,
  seed: 42,
  seat: 'north',
  maxTicks: 1_000,
});

let tick = env.reset();
while (!tick.done) {
  const action = await chooseAction(tick.observation, tick.legalActions);
  tick = env.step(action);
}

Each tick exposes the product observation, fully parameterized legal actions, reward, metrics, and separate termination and safety-truncation state. Seat-scoped environments use reducer.viewFor and never enumerate actions from the privileged full view.

The product owns observations, legal-action meaning, reward policy, and agent decision cadence. GAOS owns the environment lifecycle, concrete-action validation, deterministic transcript, and safety bounds.

Multi-agent episodes

MultiAgentEnvironment runs independent seat policies against one reducer. Sequential play applies the active seat's action; simultaneous play collects one intent per seat and advances once in canonical seat order.

Each policy receives only its seat's redacted observation. The transcript stores canonical action batches, per-seat views, rewards, and outcomes without exposing another seat's hidden state.

Drivers and tools

The provider-neutral tool adapter exposes:

• observe
• act
• reset
• transcript

The ./agent entry point includes a driver registry and keyed HTTP drivers. The Node-only ./agent-cli entry point launches MCP-capable CLIs such as Claude Code, Codex, Cursor, Grok, OpenCode, or Ollama-backed Claude Code.

gaos-agent drivers
gaos-agent status codex

gaos-agent run openai \
  --module ./environment.mjs \
  --model your-model-id \
  --seed 42

gaos-agent spawn codex \
  --mcp-url http://127.0.0.1:9000/mcp \
  --prompt "Complete the episode" \
  --tools observe,act

Model and CLI integrations must return a concrete legal SubmittedAction before the reducer sees it. Product code retains prompts, provider selection, authentication, takeover policy, and hosted MCP servers.

From evaluation to evidence

An evaluation transcript is not automatically independently verifiable. For a published result, run the driver as an ordinary signed session seat and package the canonical submissions as gaos.replay.

The verifier then needs:

1. the replay artifact;
2. the historical product reducer;
3. the semantic mappings from signed commands to reducer actions; and
4. an independently trusted identity for that verifier code.

In v0.25, the scoring authority can supply a pinned adapter directly or resolve a product-owned verifier kit containing the reducer and semantic adapter. GAOS handles the package contract, integrity, resolution, caching, and restricted execution; products retain publication, retention, identity, and score-adoption policy.

Python

Python provides a product-neutral SessionClient plus agent-evaluation helpers that accept any duck-typed Gymnasium-style environment. Typed observations and environment adapters live with each product.

Build a reducer → · Read the Python boundary → · Understand verification →