无论如何，我为你的坚持和自信点赞。

这次 Manus 发布的几项更新，我还没有体验，但 Studio 和 Cue 分别让我产生了两个问题。

首先是 Studio。我最近一直在思考，Agent 执行效果的一个重要障碍，可能是大量现实环境和隐性知识无法被 AI 高效获取。很多任务涉及的背景、依赖、经验和个人偏好，没有或无法被完整记录下来，甚至用户自己也经常无法准确表达。很多时候，人很难在事前就把要求和目标都说清楚，经常是看到结果、原型、交付物以后，才能判断是不是自己想要的，需要改成什么样。这就是我们中国人常说的“只可意会，不可言传”。
对于非程序化、非标准化的任务，这个问题更加明显。
所以 Studio 允许用户通过直接编辑和持续反馈表达意图，我觉得是一个值得关注的方向。也很好奇，它是否能够通过观察用户的修改过程，逐渐理解和修正对用户意图、目标、评价标准和价值取舍的判断？实际能在多大程度上改善这个问题？

其次是 Cue，它让我产生了另一个疑问。
如果多个 Agent 拥有独立身份和执行能力，又都作为同一个用户的受托代理参与协作，它们如何保持对用户整体意图的一致理解？
首先，用户自己就很难精确判断每个 Agent 应该获得多少上下文、什么颗粒度，以及这些信息与目标的相关性。退一步说，即使 Agent 能够自行获取相关上下文、过滤冗余信息，也不意味着基于不同局部上下文形成的决策能够自然合并。
每个 Agent 的局部判断和执行都可能是合理的，但组合起来却可能发生目标内部、目标与执行路径，以及不同局部决策之间的冲突。更复杂的是，一个 Agent 的行动可能改变其他 Agent 所依赖的条件，进而引发连锁变化。对于非标准化任务，这种情况可能相当普遍。
它暴露了一个在我看来长期被混淆的问题：分工和协作不是一回事。分工要求任务具有一定的可拆解性，协作则要求各个局部的判断和行动能够服从整体目标。
所以我很好奇 Cue 是怎么处理的：是否存在一套共享的上下文和协调机制，用来维持用户整体意图的一致性？出现冲突以后，哪个 Agent 有权代表用户决定取舍？又在什么情况下需要把决定权交还给用户？
我怀疑，这个问题不能仅仅依靠给每个 Agent 提供更完整的上下文来解决。

In any case, I respect your conviction in backing Manus again.

I haven't had a chance to try the new Manus releases yet, but Studio and Cue have raised two questions for me.

First, Studio.

One problem I've been thinking about lately is how much of the context and tacit knowledge a task depends on remains inaccessible to AI. A lot of the background, dependencies, experience and personal preferences involved in real-world tasks aren't documented, or simply can't be captured in full. Sometimes even users themselves struggle to articulate what they actually want.

Quite often, it's only after seeing a result, a prototype or a deliverable that people can tell whether it's what they had in mind, and what needs to change. There's a Chinese saying, "只可意会，不可言传" — something you can intuitively grasp but struggle to put into words.

This becomes particularly apparent with less procedural, non-standardised tasks.

That's why I find Studio's approach interesting — allowing users to express their intent through direct editing and continuous feedback. I'm curious whether Studio can also learn from those edits and gradually refine its understanding of the user's intent, goals, evaluation criteria and trade-offs. How far can that go in practice?

Cue raises another question for me.

If multiple agents have their own identities and can act independently, while all acting on behalf of the same user, how do they maintain a consistent understanding of that user's overall intent?

To begin with, users themselves often struggle to determine how much context each agent should receive, at what level of detail, or how relevant each piece of information might be to the goal. Even if agents can retrieve context on their own and filter out irrelevant information, that doesn't mean decisions made from different partial views of the context will necessarily fit together.

Each agent's local judgement and actions might be perfectly reasonable, yet the combined result could still contain conflicts — between sub-goals, between goals and execution paths, or between different local decisions. It gets more complicated when one agent's actions change the conditions that other agents are working under, potentially triggering a chain of further changes. I suspect this could be quite common in non-standardised tasks.

To me, this exposes a distinction that often gets overlooked: division of labour and collaboration are not the same thing. Division of labour requires a task to be decomposable to some extent. Collaboration requires those separate decisions and actions to remain aligned with the overall goal.

So I'm curious how Cue handles this. Is there a shared context and coordination mechanism that keeps the agents aligned with the user's overall intent? When conflicts arise, which agent has the authority to make trade-offs on the user's behalf? And when should that decision be handed back to the user?

I suspect simply giving every agent more complete context won't be enough to solve this.

These are interesting alignment questions - at the end of the day i think the game is mining both the stock and flow of the user context, meaning existing context plus probing questions to try to structure the intent. At the same time, i’ve had friends tell me that they will spend a ton of time in voice mode to build a richer PRD before getting started on agentic work. For example going for a walk for 20 minutes to explore an idea with ai before settling on a plan. For your second question around agentic context management and collaboration there has been some papers published about optimal number of agents, etc. but it seems like a lot of the constraint is again around the contextual data moat. I’m not convinced that agent teams need to be human like entities that all interact in plain sight, versus a sub agent process that just parallelizes work and aligns with the orchestrator from time to time.

Thanks for the thoughtful reply — it gave me a few new things to think about.
On understanding user intent, I agree that this approach is probably good enough if the target is simply “pretty good”. But if the target becomes “this is exactly what I wanted”, my guess is that the marginal cost of adding more context and more probing questions starts to rise quite quickly.
People naturally operate within multiple layers of context. Our worldview, values, long-term experience and preferences, and the immediate situation all shape how we judge and execute a task, often without us being fully aware of it. A lot of that context has never been properly recorded, and some of it has never even been put into words. AI only gets access to a fraction of it, so the closer we want the output to match what a person actually wants, the more information has to be surfaced and calibrated through interaction. That interaction cost can get high. For high-value, complex tasks with clear evaluation criteria and reusable context, the ROI is probably much better.
On multi-agent systems, I can see why having one orchestrator managing multiple sub-agents is a workable direction. It gives the system a clear coordination and arbitration point. But I’m not sure it fully solves the deeper issue: sub-tasks are coupled both to the overall objective and to each other, and those relationships can change as execution progresses. A result from one critical sub-task may change the overall goal, constraints, the importance of other tasks, or even invalidate work that has already been completed, which can then trigger further changes elsewhere.
So my current view is that the hardest part of multi-agent systems is not just dividing the work or syncing context. It is maintaining consistency between what is locally reasonable and what is still globally reasonable as the task evolves, without letting the objective drift without the user noticing.

感谢你的解答！让我有了新的思考。
关于精准识别用户意图，如果目标只是做到“还不错”，我同意这个方案在当前技术条件下是可行的。但如果目标变成“太棒了，这就是我想要的”，继续增加上下文和 probing questions，我猜边际成本会越来越高。
人天然生活在多层上下文里，我们的世界观、价值观、长期经验和偏好，以及当前情境，都会显性或隐性地参与任务的判断和执行，其中很多从来没有被完整记录，甚至没有被语言化。AI能够获得的只是其中一部分，因此越想符合人的需求，就越需要持续从人这里获取和校准这些信息，交互成本也随之上升。对于价值高、评价标准明确，而且上下文能够复用的复杂任务，这个 ROI 可能会好很多。
关于多 Agent，也许由一个父 Agent 统管多个子 Agent 确实是一个可行方向，它提供了统一的协调和仲裁角色。但我不确定它能否解决更深一层的问题：子任务和整体目标之间、不同子任务之间存在耦合，而且这些关系会随着执行结果动态变化。某个关键子任务产生的新结果，可能反过来改变整体目标、约束、其他任务的权重，甚至已经完成任务的有效性，并继续产生连锁变化。
所以我认为，多 Agent 最难的是如何在执行过程中持续维持“局部合理”和“整体合理”之间的一致性，并避免未经确认的目标漂移。

@PatrickKavanagh Just in case the English version got buried under the Chinese in my last reply — it’s right below it 🙂
Curious what you think about the Studio and Cue questions. The Cue one especially keeps nagging at me: if several agents are all acting on behalf of the same user, how do you stop individually reasonable decisions from adding up to something that no longer makes sense as a whole?

@PatrickKavanagh 

今年夏天，Manus 从 Meta 重新独立出来，并由原有投资人回购。Chris 和我当时面临一个选择：是否重新参与进来，并把出售所得再投回这家公司。我们答应了，也为此感到无比兴奋。智能体时代才刚刚开局，而 Manus 是我们见过产品执行力最强的团队之一。
我们很高兴看到他们今天发布了多项重大新产品，包括 Cue、Manus 2.0 和 Manus Studio——这些都站在智能体工作的最前沿。

This summer, Manus spun back out of Meta and was repurchased by its existing investors. Chris and I had a choice to get back involved and put our sale proceeds back into the company. We agreed and couldn’t be more excited. We are in the early innings of an agentic future, and Manus is one of the greatest product execution teams we’ve ever met.

We’re excited to see them announce major new products today, including Cue, Manus 2.0, and Manus Studio - all of which sit at the cutting age of agentic work.

旧金山

Old Friendship 创业投资人 ChatCut, Wanaka, Manus, Plaud, Robinhood

100万+名即友在侃科技圈那些事儿

来自圈子

科技圈大小事