EricSun0218 · EricSun0218 · Aug 13, 2026 · Aug 13, 2026
@@ -2,16 +2,42 @@
 
 [简体中文](README.zh-CN.md)
 
-**The agent kernel for games.** A compact, hackable C# runtime for building AI-native games, autonomous NPCs, and interactive worlds in Godot, Unity, or .NET services.
+**Pi for games. Give every NPC a Codex-style action loop, not just ChatGPT-style dialogue.**
+
+OpenGameAgent is a compact, hackable C# agent runtime for AI-native games, autonomous NPCs, and interactive worlds in Godot, Unity, or .NET services. An NPC can observe structured game state, plan, call game tools, inspect authoritative results, remember, and continue working—not merely generate the next line of dialogue.
 
 [![CI](https://github.com/EricSun0218/OpenGameAgent/actions/workflows/ci.yml/badge.svg)](https://github.com/EricSun0218/OpenGameAgent/actions/workflows/ci.yml)
 [![License](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](LICENSE)
 [![Status](https://img.shields.io/badge/status-alpha-orange.svg)](CHANGELOG.md)
 
-OpenGameAgent brings the small, composable agent-kernel model to game development. Its stateful core streams model output, executes validated tools, accepts steering while running, and continues the model/tool loop until work is complete. Use that kernel by itself, add the game layer for game time and durable state, then opt into extension packages for memory, goals, host-verified task plans, artifacts, delegation, external tools, structured interaction, and workflow graphs.
+Like Pi, OpenGameAgent starts with a small, composable agent kernel. It brings that model to game development: the stateful core streams model output, executes validated tools, accepts steering while running, and continues the model/tool loop until work is complete. Use that kernel by itself, add the game layer for game time and durable state, then opt into extension packages for memory, goals, host-verified task plans, artifacts, delegation, external tools, structured interaction, and workflow graphs.
 
 Inputs are bounded JSON. They may represent dialogue, combat observations, simulation ticks, UI events, plans, sensor state, or any other game-owned data; natural language is optional. No model is bundled. Cloud and local API endpoints are both supported.
 
+## What “Pi for games” means
+
+[Pi](https://github.com/earendil-works/pi) gives developers a small, programmable agent foundation rather than a fixed AI product: choose a model, register tools, compose extensions, stream events, steer an active run, and let the model/tool loop continue until the task is finished. OpenGameAgent follows that same developer model in C# and applies it to games.
+
+The kernel remains useful on its own as a general agent loop. The optional game runtime adds the coordinates that an interactive simulation needs: sessions and actors, game timelines and ticks, structured world context, bounded multi-NPC concurrency, persistent memory and tasks, engine-thread handoff, and durable action receipts. These capabilities are composed through typed interfaces instead of being hard-coded into one genre or world model.
+
+## Every NPC can work like Codex, not just chat like ChatGPT
+
+A chatbot-style NPC receives a prompt and returns dialogue. A Codex-style NPC receives a goal and the current environment, chooses tools, performs work, observes the result, and continues until it reaches a terminal outcome. In OpenGameAgent, those tools are ordinary game-owned operations: move, inspect, trade, build, schedule, recruit, investigate, or any other capability the developer exposes.
+
+| Chatbot-style NPC | OpenGameAgent NPC |
+| --- | --- |
+| Reads the latest dialogue | Observes bounded JSON containing dialogue, world state, events, UI input, sensor data, or simulation ticks |
+| Produces the next line of text | Streams text and typed tool calls, then consumes structured tool results |
+| Stops after one model response | Can observe → decide → act → inspect the result → continue across multiple turns |
+| Treats generated text as the outcome | Requests actions while game code validates permissions, rules, revisions, and state changes |
+| Usually follows wall-clock chat history | Can reason against game time, timelines, save/session identity, actor identity, and scoped memory |
+| Models one conversation at a time | Serializes each actor while allowing bounded concurrency across many NPCs |
+| May repeat a write after a timeout | Can journal state-changing intents and reconcile authoritative receipts before retrying |
+
+This does **not** mean the project bundles or depends on Codex, ChatGPT, or a particular model. The names describe the interaction pattern: an NPC is an agent operating through developer-defined tools, not a language model directly controlling game state. The game remains authoritative at every mutation boundary.
+
+Not every interaction needs the full loop. A game can route greetings and other simple inputs through the quick-response path, use the complete agent loop for open-ended tasks, and use deterministic workflows where the execution graph should be fixed.
+
 > Current version: `0.3.0-alpha.2`. Public APIs can change before `1.0`.
 
 The kernel boundary is intentionally small and designed to stabilize early. New game-specific capabilities should normally arrive as extensions, tools, policies, workflows, or game-owned services instead of expanding the model/tool loop.

@@ -2,16 +2,42 @@
 
 [English](README.md)
 
-**为游戏而生的 Agent 内核。** 一个紧凑、可修改的 C# Runtime，用于在 Godot、Unity 或 .NET 服务端构建 AI 原生游戏、自主 NPC 与互动世界。
+**游戏版 Pi：让每个 NPC 都拥有类似 Codex 的行动循环，而不只是像 ChatGPT 一样聊天。**
+
+OpenGameAgent 是一个紧凑、可修改的 C# Agent Runtime，用于在 Godot、Unity 或 .NET 服务端构建 AI 原生游戏、自主 NPC 与互动世界。NPC 可以观察结构化游戏状态、制定计划、调用游戏工具、检查权威执行结果、形成记忆并继续完成任务，而不只是生成下一句对话。
 
 [![CI](https://github.com/EricSun0218/OpenGameAgent/actions/workflows/ci.yml/badge.svg)](https://github.com/EricSun0218/OpenGameAgent/actions/workflows/ci.yml)
 [![License](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](LICENSE)
 [![Status](https://img.shields.io/badge/status-alpha-orange.svg)](CHANGELOG.md)
 
-OpenGameAgent 把小型、可组合的 Agent 内核带进游戏开发。它的有状态核心会流式接收模型输出、执行经过校验的工具、在运行中接受 steering，并持续进行模型/工具循环直到任务结束。开发者既可以只使用这个内核，也可以叠加游戏层获得游戏时间与可靠状态，再按需加入记忆、目标、宿主证据校验的任务清单、产物、委派、外部工具、结构化交互和工作流图等扩展。
+与 Pi 一样，OpenGameAgent 从小型、可组合的 Agent 内核出发，并把这种模式带进游戏开发。它的有状态核心会流式接收模型输出、执行经过校验的工具、在运行中接受 steering，并持续进行模型/工具循环直到任务结束。开发者既可以只使用这个内核，也可以叠加游戏层获得游戏时间与可靠状态，再按需加入记忆、目标、宿主证据校验的任务清单、产物、委派、外部工具、结构化交互和工作流图等扩展。
 
 输入是有大小限制的 JSON，可以表示对话、战斗观察、模拟 Tick、UI 事件、计划、传感状态或任意游戏数据，不要求是自然语言。项目不捆绑模型，同时支持云端和本地 API。
 
+## “游戏版 Pi”是什么意思
+
+[Pi](https://github.com/earendil-works/pi) 提供的是一个小型、可编程的 Agent 基础，而不是固定形态的 AI 产品：开发者可以选择模型、注册工具、组合扩展、订阅流式事件、在运行中 steering，并让模型/工具循环持续执行到任务结束。OpenGameAgent 在 C# 中采用同样的开发方式，并把它带进游戏。
+
+内核本身可以作为通用 Agent Loop 单独使用。可选的游戏 Runtime 则增加互动模拟真正需要的坐标：会话与角色、游戏时间线与 Tick、结构化世界上下文、多 NPC 有界并发、持久记忆与任务、引擎主线程交接，以及可恢复的动作回执。这些能力通过类型化接口组合，不绑定某一种玩法、题材或世界数据模型。
+
+## 每个 NPC 都可以像 Codex 一样做事，而不只是像 ChatGPT 一样聊天
+
+聊天式 NPC 接收一段提示词，然后返回一句对话。Codex 式 NPC 接收目标和当前环境，选择工具、执行工作、观察结果，并持续行动直到得到明确结果。在 OpenGameAgent 中，这些工具就是普通的游戏业务能力：移动、观察、交易、建造、安排日程、招募、调查，或开发者允许 NPC 使用的任何操作。
+
+| 聊天式 NPC | OpenGameAgent NPC |
+| --- | --- |
+| 主要读取最近的对话 | 观察有界 JSON，其中可以包含对话、世界状态、事件、UI 输入、传感数据或模拟 Tick |
+| 生成下一句文本 | 流式生成文本和类型化工具调用，并读取结构化工具结果 |
+| 一次模型回复后结束 | 可以“观察 → 决策 → 行动 → 检查结果 → 继续”，跨多轮完成任务 |
+| 把生成文本本身当作结果 | 只提出动作请求，由游戏代码校验权限、规则、版本和状态变更 |
+| 通常依赖现实时间与聊天记录 | 可以使用游戏时间、时间线、存档/会话身份、角色身份和作用域记忆 |
+| 一次处理一段对话 | 同一角色串行，不同 NPC 之间有界并发 |
+| 超时后重试可能重复写入 | 可以先记录状态变更意图，并在重试前核对游戏返回的权威回执 |
+
+这**不代表**项目捆绑或依赖 Codex、ChatGPT 或某个指定模型。这两个名字描述的是交互方式：NPC 是一个通过开发者定义的工具做事的 Agent，而不是一个直接控制游戏状态的大模型。所有状态变更始终由游戏裁决。
+
+也不是每次互动都必须运行完整 Agent Loop。问候等简单输入可以走快速回复路由，开放式任务使用完整循环，需要固定执行图的场景则可以使用确定性 Workflow。
+
 > 当前版本：`0.3.0-alpha.2`。在 `1.0` 前公开 API 仍可能调整。
 
 内核边界刻意保持小而稳定。后续游戏特有能力通常应通过扩展、工具、策略、工作流或游戏自有服务加入，而不是继续膨胀模型/工具循环。