文章背景与核心概要

生成式智能体系统（Generative-agent systems）历来以难以审查而著称。虽然单次运行可能涉及无数的智能体、位置、消息、指令和底层模型调用，但操作员通常只能面对静态的完成回放或繁琐的原始日志。这种透明度的缺乏使得回答基本问题（例如：“为什么特定 agent 会移动到那里？”）、测试微小干预或打包仿真运行以实现可重复性变得极其困难。

为了解决这一痛点，作者推出了 GOD（Govern, Observe, and Direct，治理、观察与指挥），这是一个专门为智能体社会设计的高效、本地优先的控制室。通过统一的浏览器工作流，GOD 赋能操作员能够：发出有针对性的问题和实时干预；动态检查实时回放状态；利用包括设置向导、Agent Studio、Map Studio、空间回放界面和交互式命令界面在内的模块化工具；并通过运行时状态与领域数据之间的清晰分离契约，管理可移植的场景、地图和智能体包。该系统已被 EMNLP 2026 系统演示轨道（System Demonstrations Track）接受。

GOD: Govern, Observe, and Direct — A Real-Time Control Room for Agent Societies

Authors: Yige Luo, Ran Guan
Published: arXiv:2608.27992 [cs.AI] — August 28, 2026
Status: Accepted to the EMNLP 2026 System Demonstrations Track (9 pages, 5 figures)

📌 Executive Summary

Generative-agent systems are notoriously difficult to inspect. While a single run can involve countless agents, locations, messages, commands, and underlying model calls, operators are typically left with either static finished replays or cumbersome raw logs. This lack of transparency makes it exceptionally difficult to answer basic questions (e.g., "Why did a specific agent move there?"), test minor interventions, or package a simulation run for reproducibility.

To solve this, the authors introduce GOD (Govern, Observe, and Direct), a local-first control room designed specifically for agent societies. Through a unified browser workflow, GOD empowers operators to: * Issue targeted questions and live interventions. * Inspect real-time replay states dynamically. * Utilize modular tools including a setup wizard, Agent Studio, Map Studio, a spatial replay interface, and interactive command interfaces. * Manage portable scenarios, maps, and agent packages via clean separation contracts between runtime state and domain data.

🛠️ System Architecture & Components

GOD 结合了几个核心模块，以弥合执行与可观测性之间的鸿沟：

GOD combines several core modules to bridge the gap between execution and observability:

设置向导与工作室： 简化智能体环境的初始化。
Agent Studio： 管理智能体配置文件和行为属性。
Map Studio： 配置空间环境和导航布局。
1. Setup Wizard & Studios: Simplifies the initialization of agent environments.
2. Agent Studio: Manages agent profiles and behavioral attributes.
3. Map Studio: Configures spatial environments and navigational layouts.
空间回放界面： 提供实时可视化窗口，用于观察实时展开的智能体社会。
1. Spatial Replay Interface: Offers a live visualization window to observe agent societies unfolding in real time.
命令与构件循环： 一个统一的操作员命令模型，其中实时控制和回放证据共享相同的底层架构。
1. Command & Artifact Loop: A unified operator command model where live controls and replay evidence share the same foundational architecture.
可移植包： 通过严格的包契约，场景、地图和智能体数据与本地运行时状态彻底解耦，从而使实验易于共享和复现。
1. Portable Packages: Scenario, map, and agent data are cleanly decoupled from local runtime state via strict package contracts, making experiments easy to share and reproduce.

📊 Evaluation & Results

该系统在 15 个已完成的运行槽（run slots） 中进行了严格评估： * 干预准确率： 在 14 次干预运行中，84 个目标智能体检查中有 78 个成功记录了被指挥的目的地。 * 状态一致性： 182 个状态回答中有 169 个准确匹配了预期的保存位置或动作字符串。 * 公共资源： 公开发布的内容包括托管的 Smallville 风格和 PKU 回放、开源代码库以及可下载的配置包。

The system was rigorously evaluated across 15 completed run slots: * Intervention Accuracy: Across 14 intervention runs, 78 out of 84 target-agent checks successfully recorded the commanded destination. * State Consistency: 169 out of 182 state answers accurately matched the expected saved location or action string. * Public Resources: The public release includes hosted Smallville-style and PKU replays, an open-source code repository, and downloadable configuration packs.