How seenzus thinks seenzus 如何思考

Designing Agents for Real-World Spaces: Tools, Execution, and Memory at seenzus为真实空间设计 Agent：seenzus 的工具、执行与记忆

A light-control request shows how seenzus prepares device parameters and explains execution results. We look at which model calls can be skipped, and how to remember a condition such as “only tonight.”我们从调灯说起，讲 seenzus 怎样把具体设备的参数准备好，按证据解释执行结果。再看哪些模型调用可以省，以及用户说的“只在今晚”该怎样记住。

Does seenzus get better as models get stronger?模型越强，seenzus 就越好吗？

GPT-6 raises the ceiling; a physical-world product determines how much reaches realityGPT‑6 抬高能力上限，物理世界产品决定它能落下多少

Model capability matters greatly to seenzus, but the relationship is not simple substitution. The model determines how much ambiguity and change a Space Agent can handle; seenzus supplies persistent spatial context, memory, authority, and action so that general intelligence becomes useful to a place.模型强弱与 seenzus 高度相关，却不是简单替代关系。模型决定 Space Agent 能处理多少含糊与变化；seenzus 提供持续的空间上下文、记忆、权限和行动，让通用智能真正对一个地方有用。

Anthropic's MHS gives us more confidence in physical agentsAnthropic 的 MHS，让我们更确信物理 Agent 的方向

Before MHS was announced, seenzus had already chosen to put agents into real spaces. Now two independent paths point toward the same physical agent direction.在 MHS 发布前，seenzus 已经选择让 Agent 进入真实空间。如今，两条独立路径指向同一个物理 Agent 方向。

Physical agents were already part of the seenzus product direction. Anthropic's MHS adds evidence from the infrastructure side and gives us more confidence that this path deserves long-term investment.seenzus 早已把物理 Agent 写进产品方向。Anthropic 的 MHS 从行业基础设施一侧提供了新的证据，让我们更确信这条路值得长期投入。

A good Space Agent can actively uncover what users need好的 Space Agent 可以主动挖掘出用户的需求

Many conversational products understand natural language, yet users still have to know what the product can do before they can ask the right question. seenzus lets the Space Agent combine the live conversation, personal usage history, and the state of the space to uncover an unstated need and offer a relevant next step.许多对话产品已经能听懂自然语言，用户仍要先知道产品会做什么，才能问出那句正确的话。seenzus 让 Space Agent 结合当前对话、个人使用历史和空间现状，主动发现一个尚未明确提出的需求，再给出贴合当下的下一步。

Why we call it a World我们为什么把它叫作「世界」

Places belong to one World when they share an Agent, environmental memory, and Rules. A building type describes a place; it cannot choose that boundary for its owner.多个地点是否属于同一个 World，取决于它们是否共享 Agent、环境记忆和 Rules。房屋类型只说明地点是什么，不能替人划这条边界。

The next layer of agent context lies beyond the screenAgent 的下一层上下文，在屏幕之外

Today's models are good at reading conversations, documents, code, and web pages. seenzus is working on another layer: helping an agent understand its places, spaces, devices, and the changes that keep happening outside conversation.今天的模型很擅长读取对话、文档、代码和网页。seenzus 想补上另一部分：让 Agent 持续理解它所处的地点、空间、设备，以及那些在对话之外不断发生的变化。

We took learning back from the machinery we built for it我们把学习从自己搭的机器手里要了回来

In July, learning was a seven-station pipeline and the agent was the station that stamped. In August it shrank to one review. Every machine had fixed a real problem, and together they took judgment out of the agent's hands. Once the model got strong, the parts that judged in its place came down.七月，学习还是一条七站流水线，Agent 只是负责盖章的那一站。八月，它收缩成一次审阅。那些机器当初每一台都在修真问题，也一件件把判断从 Agent 手里拿走了。模型变强以后，该拆的是替模型做判断的部分。

Learning a habit is not permission to act学到一个习惯，不等于获得自动执行权

seenzus moves a habit through observation, asking, and automatic care. Permission is earned for one exact action; timing may adapt, while the device and parameters may not drift.seenzus 把习惯拆成仅观察、到点询问和自动照看。权限按具体动作逐步获得，时间可以适应，设备和参数不能随学习漂移。

What it means to understand you better over time住得越久越懂你到底意味着什么

Remembering habits is easy. Understanding change is harder.记住习惯容易，理解变化难。

Long-term memory is useful only when the user's current expression comes first. seenzus treats forgetting as stopping an old inference from shaping current service, not simply deleting history.长期记忆只有在当前表达优先时才真正有用。seenzus 把遗忘定义为停止让旧推断影响当前服务，而不是简单删除历史。

Why seenzus had to learn when not to speak管家为什么必须先学会不说话

The easiest way for a proactive agent to fail is not missing things but speaking too much. seenzus treats speaking as a scarce act applied for with evidence: one voice at a time, corrections first, uncertainty said in words.主动服务最容易失败的地方不是看漏，而是开口太多。seenzus 把开口做成凭证据申请的稀缺动作：同一时刻一个声音，被否定的先纠正，说不准就用话承认。

Why agent memory is starting to look like a set of filesAgent 的记忆为什么越来越像一套文件

We examined knowledge-base RAG, CLAUDE.md, progressive disclosure, and the memory systems in Claude Code, Codex, and Hermes. The harder questions are what survives, what enters the next turn, and where the boundary belongs in a shared physical space.我们回头看了知识库 RAG、CLAUDE.md 和渐进式披露，又拆开 Claude Code、Codex 与 Hermes 的实现。除了存储方式，我们更关心记忆怎样留下、怎样进入下一轮，以及到了多人空间里，边界该画在哪里。

Why we did not build another wall of switches我们为什么不想再做一面智能家居开关墙

A conventional smart home waits for someone to inspect devices and issue commands. seenzus is a Smart Space Agent that involves the user when attention or a decision is required.传统智能家居等待用户检查设备并发出指令。seenzus 选择成为 Smart Space Agent，只在需要用户关注或决定时提出。