模块 1 · How Muse Works
第 2 节,共 5 节 · 12 分钟 · 英语 · 只读
This lesson traces one request through Muse's main stages so you can see it as a pipeline rather than a single black box. You will meet the system boundary, the input stage, the core agentic model, and the output stage, then follow a booking request through them in order.
Muse is a personal AI agent: it does not just answer questions, it takes action on a person's behalf. The sources describe it as running on Muse Secure VM, a dedicated cloud computer with its own browser, where the agent and the person's data live together, and it can work across the apps a person uses daily. That framing matters for this lesson because it tells you the thing you are tracing is a system with parts, not one monolithic model call. A useful mental model is a pipeline: something comes in, a core agent decides and acts, and something observable comes back out. The next blocks make each stage of that pipeline inspectable.
来源支持
A request moves in one direction: a stated goal is processed into a plan, the core agent acts on it, and results are generated and reported. Approval checkpoints and background work are described in the prose, not encoded as extra edges here.
来源支持
Input processing is where a stated goal becomes something actionable. The sources say that once a person shares a goal, Muse helps develop a personalized plan and coordinate their time and resources, then advances the work on its own. The core agentic model is Muse Spark, described as a multimodal reasoning model built for agentic tasks; as the main agent it can gather context, make a plan, and delegate execution across parallel subagents, and it can open a browser and fill out forms. Output generation is the visible result: tasks completed, plus a complete audit trail of everything Muse has done and plans to do. Two boundaries belong to this picture rather than to a separate stage. First, Muse checks with the person before sensitive actions like sending an email or making a purchase, so the pipeline can pause for approval. Second, for tasks that take more time, Muse keeps working after people close the app and comes back when something changes or when it needs approval.
来源支持
The stages are not just a convenient diagram; they correspond to different things happening in the system. Input processing is where a goal becomes a plan, which is why the same goal can produce different plans for different people. The core agentic model is where reasoning and tool use happen, which is why the same plan can be executed in different ways depending on what the agent finds. Output generation is where results and the audit trail appear, which is why you can inspect what happened even when the agent worked in the background. Treating Muse as one monolithic step hides all three of these, and it makes it hard to say where a problem occurred when something goes wrong.
来源支持
想一想
Pick one task you would hand to an agent. Which stage of the pipeline would you want to inspect most closely before trusting the result, and what would you look for there?
A person tells Muse they want to book travel for a trip. Using only the stages and behaviors described in this lesson, trace what happens at each stage. Assume the person has already connected the relevant apps and granted the access Muse needs.
A stated goal — book travel — arrives through the conversation-style interface. No plan exists yet, and no action has been taken.
Receive the goal as input.
为什么是这一步
Muse is designed around the way people already communicate, so a goal arrives as ordinary conversation rather than a technical command.
发生了什么变化
The goal is now inside the system, but it is still just a request; nothing has been planned or executed.
来源支持
Turn the goal into a personalized plan and coordinate time and resources.
为什么是这一步
The sources state that once a person shares a goal, Muse helps develop a personalized plan and coordinate their time and resources before advancing the work.
发生了什么变化
The request now has structure: an ordered plan with the resources it needs, rather than a single undifferentiated instruction.
来源支持
Advance the work using the core agentic model: gather context, act with tools, and open a browser to fill out forms.
为什么是这一步
Muse Spark is described as built for agentic tasks and able to open a browser, fill out forms, and negotiate on the person's behalf.
发生了什么变化
Concrete progress is made on the booking rather than only a description of what could be done.
来源支持
Pause for approval before the sensitive action, then produce the result and the audit trail.
为什么是这一步
Muse checks with the person before sensitive actions like making a purchase, and it shows a complete audit trail of everything it has done and plans to do.
发生了什么变化
The pipeline does not run straight through to completion on its own authority; it stops at the approval point, and the person can see what happened and what is planned.
来源支持
The request moves through four distinguishable stages — goal input, input processing into a plan, action by the core agentic model, and output generation with an approval checkpoint — rather than being handled in one opaque step. The booking is advanced by the agent, but the purchase itself waits for the person's approval, and the person can review an audit trail of what was done and planned.
来源支持