Name: 人工智能代理的下一個突破就在這裡 (The Next Breakthrough In AI Agents Is Here)
Uploaded: 2025-04-12T12:12:51.000Z
Duration: 8 min 30 s

Usable AI agents are finally here, from deep research platforms out of OpenAI and Google, to similar tools from XAI and DeepSeek.

從 OpenAI 和谷歌的深度研究平臺，到 XAI 和 DeepSeek 的類似工具，可用的人工智能代理終於出現了。

關係子句

Joining the competition now is Manus, a brand new agentic AI platform that has taken the world by storm.

現在加入競爭行列的是 Manus，這是一個全新的代理人工智能平臺，已風靡全球。

And today, we're launching an early preview of Manus, the first general AI agent.

今天，我們推出了首個通用人工智能代理 Manus 的早期預覽版。

When Manus officially launched, the hype around it immediately took off.

Manus 正式推出後，立即引起了熱議。

A Chinese startup unveiling a new AI agent that some are calling China's next DeepSeek With people calling it the most impressive AI tool they've ever tried and the most sophisticated computer using AI.

一家中國初創公司推出了一款新的人工智能代理，有人稱其為中國的下一個 DeepSeek，還有人稱其為他們嘗試過的最令人印象深刻的人工智能工具和使用人工智能的最複雜的計算機。

Unlike some of its predecessors, Manus wasn't just another specialized chatbot.

與前幾款哈拉機器人不同，Manus 不只是另一款專門的哈拉機器人。

It promised to be a true general purpose AI agent.

它承諾成為一個真正的通用人工智能代理。

With invitations rare and access limited, the question remains, has Manus truly revolutionized the AI agent landscape?

在邀請稀少、訪問受限的情況下，問題依然存在：Manus 是否真正徹底改變了人工智能代理的格局？

Behind all the excitement around Manus is something genuinely innovative, a multi-agent AI system that can seemingly complete all sorts of tasks, from travel planning and financial analysis to searching over dozens of files or doing industry research.

在圍繞 Manus 的所有激動人心的事件背後，是一種真正的創新--多代理人工智能系統，它似乎可以完成各種任務，從旅行規劃和財務分析到搜索數十個文件或進行行業研究。

Rather than relying on one big neural network, Manus works more like an executive overseeing a team of sub-agents coordinating and guiding their every move across a shared action space.

與其說 Manus 依靠的是一個龐大的神經網絡，不如說它更像是一個管理者，負責監督一個由子代理組成的團隊，協調和指導他們在共享行動空間內的一舉一動。

It takes in your prompt as input and gets to work figuring out what it needs to do.

它將你的提示作為輸入，然後開始工作，找出它需要做的事情。

Instead of tackling your task in one go, a planner agent first comes up with a master plan to follow, breaking things down into manageable subtasks.

計劃員不會一次性完成任務，而是首先制定一個總體計劃，將事情分解成易於管理的子任務。

This way, Manus knows precisely what needs to be done before executing and can hand off these tasks to other sub-agents.

這樣，Manus 就能準確地知道在執行之前需要做什麼，並將這些任務交給其他分代理。

These are like Manus' own in-house experts.

他們就像是馬努斯自己的內部專家。

They share the same context, but each has its own delineated domain from knowledge or memory to execution.

它們共享相同的上下文，但從知識或記憶到執行，每一個都有自己劃分的領域。

Manus can call upon an extensive suite of 29 different integrated tools.

Manus 可以調用 29 種不同的集成工具。

Whether they're automating web navigation, securely running code, or pulling important information from files, Manus' sub-agents intelligently decide which tools to use.

無論是自動化網絡導航、安全運行代碼，還是從文件中提取重要資訊，Manus 的子代理都能智能地決定使用哪些工具。

Finally, when each subtask is complete, the executor agent combines the outputs together into a final, synthesized output for the user.

最後，當每個子任務完成後，執行代理會將輸出結果合併成最終的合成輸出，提供給用戶。

Under the hood, Manus is powered by a pretty sophisticated dynamic task decomposition algorithm.

在引擎蓋下，Manus 採用了相當複雜的動態任務分解算法。

This is what enables it to autonomously break down complex instructions into clear execution paths.

這使它能夠自主地將複雜的指令分解為清晰的執行路徑。

To ensure stability, even after dozens of rounds of reasoning and tool use, the Manus team developed an original technique called Chain of Thought Injection, enabling agents to actively reflect and update plans.

為了確保穩定性，Manus 團隊開發了一種名為 "思維鏈注入"（Chain of Thought Injection）的獨創技術，使代理能夠主動反思和更新計劃，即使經過數十輪推理和工具使用後仍能保持穩定。

At its core, Manus makes use of Anthropic's CLOD 3.7 SONNET.

Manus 的核心是使用 Anthropic 的 CLOD 3.7 SONNET。

Manus also features robust cross-platform execution capabilities thanks to its seamless integration with open-source tools like YC Company browser use for advanced website interaction and Startup E2B's secure cloud sandbox environment.

Manus 還具有強大的跨平臺執行能力，這得益於它與 YC 公司瀏覽器等開源工具的無縫集成，可用於高級網站交互和 Startup E2B 的安全雲沙箱環境。

那麼，馬努斯到底能做什麼呢？

Impressively, it can take on a wide range of real-world tasks.

令人印象深刻的是，它可以承擔現實世界中的各種任務。

It excels in scenarios like creating travel itineraries, detailed financial analyses, and educational content.

它在創建旅行路線、詳細財務分析和教育內容等場景中表現出色。

While it can also assist with valuable tasks like structured database compilation, insurance policy comparisons, supplier sourcing, and even assisting with high-quality presentations.

同時，它還可以協助完成結構化數據庫編譯、保險單比較、供應商採購，甚至協助進行高質量的演示等有價值的任務。

To truly measure Manus' capabilities, we can look at Gaia, a benchmark designed to challenge AI agents on reasoning, multimodal handling, web browsing, and tool proficiency.

要真正衡量 Manus 的能力，我們可以看看 Gaia，這是一個旨在挑戰人工智能代理推理、多模態處理、網頁瀏覽和工具熟練程度的基準。

Humans typically score about 92%, whereas OpenAI's deep research, in comparison, scored about 74% at its best.

人類的得分通常約為 92%，而相比之下，OpenAI 的深度研究在最佳狀態下的得分約為 74%。

Manus smashed the state-of-the-art on Gaia, scoring 86.5%, just a few points shy of the average human.

馬努斯打破了蓋亞的最先進水準，得分率高達 86.5%，僅比普通人類差幾分。

Still, despite impressive benchmark performance, Manus has reignited a broader conversation about the nature of AI startups at the application layer, wrappers.

不過，儘管基準性能令人印象深刻，Manus 還是重新引發了一場更廣泛的對話，討論人工智能初創企業在應用層（封裝層）的本質。

Some have dismissed Manus as merely a wrapper since it stitches together existing foundational models and various tool calls.

有些人認為 Manus 只是一個包裝，因為它將現有的基礎模型和各種工具調用拼接在一起。

But this dismissal overlooks an important reality.

但這種否定忽略了一個重要的現實。

Most successful AI products today could also qualify as wrappers by this logic.

按照這種邏輯，當今大多數成功的人工智能產品也可以被稱為封裝器。

Cursor and Windsurf, for example, integrate existing LLMs alongside external APIs and developer-focused tooling, such as real-time code analysis and debugging utilities.

例如，Cursor 和 Windsurf 將現有的 LLM 與外部 API 和開發人員專用工具（如實時代碼分析和調試實用程序）集成在一起。

Domain-specific agents like Harvey combine foundational models with legal-specific tool integrations, case law retrieval, compliance checks, and document analysis.

像 Harvey 這樣的特定領域代理將基礎模型與特定法律工具集成、判例法檢索、合規性檢查和文檔分析結合起來。

Clearly, many useful applications do fit the wrapper mold.

顯然，許多有用的應用確實符合包裝機的模式。

And for many developers, it makes sense to go this route.

對許多開發人員來說，走這條路是明智的。

Kejie told us himself, from day one, they decided to work orthogonally to model development, wanting to be excited rather than threatened by each new model release.

柯捷告訴我們，從第一天起，他們就決定正交地開發模型，希望每一個新模型的發佈都能讓他們感到興奮，而不是受到威脅。

What distinguishes successful wrappers from their less effective counterparts is typically a bunch of things, intuitive UI, proprietary evals, much more careful fine tuning of foundational models, and thoughtfully designed multi-agent architectures.

成功的封裝器與效果不佳的封裝器之間的區別通常有以下幾點：直觀的用戶界面、專有的驗證、對基礎模型進行更細緻的微調，以及設計周到的多代理架構。

Manus itself illustrates these tradeoffs really well.

馬努斯本身就很好地說明了這些權衡。

On the positive side, its multi-agent orchestration helps deliver significantly lower per-task costs, around $2 a task, compared to integrated competitors like OpenAI's Deep Research.

從積極的方面看，與 OpenAI 的 Deep Research 等綜合型競爭對手相比，它的多代理協調技術有助於大幅降低每個任務的成本，約為每項任務 2 美元。

Manus also offers greater transparency in user control, letting users directly inspect, customize, or replace individual sub-agents and tool integrations.

Manus 還提高了用戶控制的透明度，讓用戶可以直接檢查、定製或替換各個子代理和工具集成。

A degree of flexibility, centralized platforms rarely match.

這種靈活性是集中式平臺所無法比擬的。

One of the coolest things Manus figured out was actually exposing the file system so you could see exactly what the agents were doing.

Manus 想出的最酷的辦法之一就是公開文件系統，這樣你就能清楚地看到代理在做什麼。

ChatGPT requires you to reprompt, and it's opaque what's happening when it's thinking.

ChatGPT 需要您重新提示，而且它在思考時發生了什麼並不清楚。

Manus is a glimpse into the future of ChatGPT desktop operating directly on your computer.

Manus 是未來直接在電腦上運行 ChatGPT 桌面的一個縮影。

And it will be cool to see how much more control you'll get when it's happening there instead of a browser.

而且，如果是在那裡而不是在瀏覽器上進行，您將獲得更多的控制權，這將是一件很酷的事情。

Coordination across specialized agents becomes increasingly difficult as tasks' scale or complexity grows.

隨著任務規模或複雜程度的增加，各專業代理之間的協調變得越來越困難。

More critically, its current advantages, UX refinements, targeted fine-tuning, thoughtful integrations are vulnerable to competitors coming along and doing that as well.

更關鍵的是，它目前的優勢、用戶體驗的完善、有針對性的微調、周到的集成都很容易受到競爭對手的影響。

These strengths and weaknesses are generally shared by wrappers.

這些優缺點一般都是包裝者所共有的。

They allow you to have really rapid deployment, iteration, and specialized UX at lower upfront costs.

它們可以讓你以較低的前期成本實現真正的快速部署、迭代和專業化用戶體驗。

But they're also vulnerable to disruptions, such as API pricing changes or provider policy shifts, which can quickly erase any of the cost benefits.

但它們也很容易受到干擾，如 API 定價變化或提供商政策的轉變，這些都會迅速抹去任何成本效益。

Ultimately, the critical challenge isn't deciding whether wrappers are viable, but identifying genuinely sustainable differentiation for your product.

歸根結底，關鍵的挑戰不是決定包裝紙是否可行，而是確定產品真正可持續的差異化。

For founders, this might mean investing early and proprietary evals that are expensive or time-consuming to replicate, embedding your workflows deeply into specific user routines to increase switching costs, or identifying integrations with platforms or datasets competitors can't easily access.

對於創始人來說，這可能意味著要儘早投資，進行昂貴或耗時的專有驗證，將您的工作流程深度嵌入特定的用戶例程以增加轉換成本，或確定與競爭對手無法輕鬆訪問的平臺或數據集的集成。

In the end, success in AI doesn't hinge on reinventing the wheel, but rather on who can stitch together the existing models into a product users genuinely love.

歸根結底，人工智能的成功並不取決於重新發明輪子，而是取決於誰能將現有模型拼接成用戶真正喜愛的產品。