# Agent2Game (A2G)

> An open protocol that lets existing games ask an AI model what their characters should do next, while the game stays authoritative: the game describes itself in a manifest, sends observations and legal actions, executes validated typed choices through its own code and falls back to its native AI on any failure. Protocol version 0.1 (draft).

Pages marked "Proposed — not implemented yet" describe planned parts. The complete documentation in one file: https://agent2game.org/llms-full.txt


## Get started

- [Introduction](https://agent2game.org/docs/introduction/): What Agent2Game is, which parts it consists of and who builds each of them.
- [Architecture](https://agent2game.org/docs/architecture/): The participants in A2G and how responsibilities are split between them.
- [Integrate a game](https://agent2game.org/guides/integrate/): The steps to connect an existing game to A2G — manifest, adapter and hooks.

## Specification

- [Overview](https://agent2game.org/specification/overview/): Transport, versioning, endpoints and message flow of A2G protocol 0.1.
- [Manifest](https://agent2game.org/specification/manifest/): How a game describes its world, scopes, profiles, entity types, observations, events, actions and caching rules.
- [POST /v1/session](https://agent2game.org/specification/session/): Start a session by announcing the game, protocol version and manifest.
- [POST /v1/context](https://agent2game.org/specification/context/): Add, replace or remove scopes and profiles during a session — for example when a level loads.
- [POST /v1/decide](https://agent2game.org/specification/decide/): Request actions for a batch of entities in one call — with scope, profiles, groups, events, cached decisions and fallback codes.
- [POST /v1/report](https://agent2game.org/specification/report/): Tell the service what happened to a decision, building per-entity memory.
- [POST /v1/telemetry](https://agent2game.org/specification/telemetry/): Periodic snapshot of what the game is actually doing.
- [Modes and fallback](https://agent2game.org/specification/fallback/): How control moves between the game's native AI, a local policy and the decision service.
- [Security](https://agent2game.org/specification/security/): Threat model, deployment profiles, authentication, decision integrity, shared-cache safety, prompt injection, abuse limits and privacy in A2G.
- [Policy format](https://agent2game.org/specification/policy-format/): The portable file format for local policies — inputs, model kinds, evaluation rules and conformance vectors that every SDK must pass.

## Design

- [Scaling to complex games](https://agent2game.org/design/scaling/): How to use A2G for large worlds with levels, rich characters, groups and many players — and why cost must grow with decisions, not with the world.
- [Local policy](https://agent2game.org/design/local-policy/): Train a small in-game model on the big model's decisions during development, then ship the game with it — no service, no network, no per-decision cost.
- [Cost](https://agent2game.org/design/cost/): How much A2G decisions cost, how to calculate it, and which levers change it.
- [Decision providers](https://agent2game.org/design/providers/): Which models can make A2G decisions — OpenAI Decisions, TypeSafe Jev and open Jev-compatible models such as Kev, Laya and decider, general LLMs and local policies — and how each is connected.
- [Deployment, load and offline](https://agent2game.org/design/deployment/): Where the decision service runs, what happens when the network or provider fails, how it handles tens of thousands of players, and when to use A2G at all.

## Implementations

- [Godot SDK](https://agent2game.org/implementations/godot-sdk/): Reference A2G client for Godot 4 — what it does, its API and where to get it.
- [Reference service](https://agent2game.org/implementations/service/): The reference A2G decision service — what it does, how to run it and its limits.

## Examples

- [Lango (Godot 4)](https://agent2game.org/examples/lango/): A complete A2G integration in an existing Godot RPG — manifest with scopes, groups, events and cache, adapter, hooks, security profile, messages on the wire and debugging tools.

## Reference

- [Glossary](https://agent2game.org/glossary/): Short definitions of A2G terms. Links to these terms show the definition on hover.
- [Roadmap](https://agent2game.org/roadmap/): What comes next for A2G, and open design questions for complex games.
- [About](https://agent2game.org/about/): Who is behind Agent2Game and why it exists.