Name: 人工智能趨勢--瞭解人工智能代理 (AI Trends - Understanding AI agents)
Uploaded: 2025-07-23T18:42:04.000Z
Duration: 4 min 27 s

Wouldn't it be great if you could give an AI system a goal, like summarize these five articles in a new document or make a travel itinerary for our family trip to Norway in December and save it in a spreadsheet and then have the AI figure out what steps to complete and how to meet that goal.

如果你能給人工智能系統一個目標，比如在新文檔中總結這五篇文章，或者為我們全家 12 月的挪威之行制定一個旅行日程表並保存在電子表格中，然後讓人工智能找出需要完成的步驟和實現目標的方法，豈不美哉。

未來進行式

And based on what we've seen so far in this space, you'll likely be dealing as much if not more with AI agents as traditional chat-based AI systems in the future.

根據我們目前在這一領域看到的情況，未來人工智能代理與傳統的哈拉式人工智能系統打交道的次數可能會一樣多，甚至更多。

AI systems like ChatGPT, Bing, BART, and Cloud all have one thing in common.

像 ChatGPT、Bing、BART 和 Cloud 這樣的人工智能系統都有一個共同點。

You make a request in the form of a prompt, they provide a human language response.

你以提示的形式提出請求，他們就會以人類語言作出迴應。

So if you have a task you want completed, for example, summarize the main takeaways from these seven articles point by point, you have to do a lot of the manual labor yourself, copy and paste each article into the system, then prompt the system, then copy the response back into whatever application you're working in.

是以，如果你要完成一項任務，例如，逐點總結這七篇文章的主要收穫，你就必須自己完成大量的手工勞動，將每篇文章複製並粘貼到系統中，然後提示系統，再將回複復制到你正在使用的應用程序中。

Anytime you have a chain of thought or chain of events type request, you the human have to manually manage that chain, not only figuring out what needs to be done, but actually doing most of the work yourself.

任何時候，只要你有一連串的想法或事件請求，你這個人就必須手動管理這個鏈條，不僅要弄清楚需要做什麼，而且實際上要自己完成大部分工作。

AI agents are applications built to eliminate some or all of that extra work.

人工智能代理是為消除部分或全部額外工作而構建的應用程序。

Here's an overall idea of how AI agents operate.

下面是人工智能代理運行的總體思路。

First, a user sends off a goal prompt to the AI agent.

首先，用戶向人工智能代理髮出目標提示。

For example, summarize these five articles.

The agent then sends a prompt to the AI saying, what tasks need to be completed for us to reach this goal?

然後，代理向人工智能發出提示：我們需要完成哪些任務才能實現這一目標？

When the AI responds, the response is stored in memory, and then the response is sent to a task evaluator and a task creator that then sends a new set of tasks up to the task queue.

當人工智能做出迴應時，會將回應存儲在內存中，然後將回應發送給任務評估器和任務創建器，再將一組新的任務發送到任務隊列中。

Then we go into a circle where the AI agent goes through each of the tasks in the queue, and every time a new task is sent in, it goes to the memory and retrieves a summary of each previous response so that we retain context within the loop.

然後，我們進入一個循環，讓人工智能代理逐一處理隊列中的任務，每當有新任務發送進來，它就會進入內存，檢索之前每次響應的摘要，這樣我們就能在循環中保留上下文。

The agent will continue going through the loop and evaluating the responses until it feels like it met the original goal, at which point it passes everything back up to the user again.

代理將繼續循環並評估響應，直到它認為自己達到了最初的目標，然後再將所有資訊傳遞給用戶。

So in simple terms, AI agents are automatic multi-prompt loops that leverage the language capabilities of large language models to simulate multi-step reasoning to reach a goal.

是以，簡單地說，人工智能代理是自動的多提示循環，它利用大型語言模型的語言能力，模擬多步驟推理，以達到目標。

What makes these systems so powerful is the addition of two key features and capabilities not currently available from AI systems.

這些系統之所以如此強大，是因為增加了兩個目前人工智能系統所不具備的關鍵特性和功能。

ChatGBT and its like have limited context windows, meaning if you provide them with a lot of content or carry on a long chat with them, they forget the earlier parts of the conversation.

ChatGBT 和類似軟件的上下文窗口是有限的，這意味著如果您向他們提供大量內容或與他們進行長時間哈拉，他們就會忘記對話的前半部分。

AI agents solve this by summarizing the response to every prompt, storing that response in local memory, and adding the summaries as context to future prompts.

人工智能代理解決這個問題的方法是總結對每個提示的迴應，將回應存儲在在地內存中，並將總結作為上下文添加到未來的提示中。

Second, they have the ability to perform actions on behalf of the user, including doing internet searches and downloads, logging into services, filling out forms, creating files, even installing and configuring software on your computer.

其次，它們有能力代表用戶執行操作，包括進行互聯網搜索和下載、登錄服務、填寫表格、創建文件，甚至在計算機上安裝和配置軟件。

There are new AI agents coming online on a weekly basis, and a lot of research is being done on how to build, improve, and expand on this general idea.

每週都有新的人工智能代理上線，人們正在就如何構建、改進和擴展這一總體思路進行大量研究。

From my perspective, AI agents are a potentially fruitful path forward, and I expect we'll see a rapid maturing of the utility of these services over the next month and years.

在我看來，人工智能代理是一條富有成果的潛在發展道路，我預計在未來一個月或幾年內，這些服務的實用性將迅速成熟起來。

Additionally, projects like Microsoft Copilot and ChatGBT plugins are learning from and leaning into the same general direction with the introduction of context memory, online access, and agent capabilities.

此外，Microsoft Copilot 和 ChatGBT 插件等項目也在學習和借鏡同樣的方向，引入了上下文記憶、在線訪問和代理功能。

In summary, AI agents are a glimpse at our possible near future, where goal-oriented AIs are able to do more than simply respond to our prompts with human-sounding language, and you'll be able to say what you want and have an AI figure out how to do it for you.

總之，人工智能代理讓我們看到了不遠的將來，在那裡，以目標為導向的人工智能不僅能用聽起來像人類的語言來回應我們的提示，還能說出你想要什麼，讓人工智能為你想辦法。