How does an AI Agent achieve true intelligence?

To be honest, when I see the AI Agent projects on the street now, I often wonder: how far are these so-called "intelligent" systems from real autonomous thinking? Recently, I have read several in-depth analyses, especially about the Manus case, which was once very popular, and it makes me re-examine this question. True intelligence may not lie in how much it can do, but in whether it can continuously learn and self-optimize in a specific scenario, and even correct its own mistakes without human intervention. This reminds me of a common saying in the field of autonomous driving: the accuracy of 99% is excellent in the laboratory, but every 100 times on the real road there will be an accident - a price that no one can afford.

The heart of intelligence: from reactive response to proactive planning

Nowadays, most of the AI Agent is still stuck in the mode of "you ask, I answer", like an advanced version of the search engine. But the real intelligence should be able to be like a human assistant, not only understand the instructions, but also actively plan the steps of the task. For example, when you let a truly intelligent Agent to help you prepare for a meeting, it should be able to automatically find relevant information, booking conference rooms, coordinating participants time, and even after the meeting to generate the minutes and follow up on the to-do list - this entire process requires dynamic planning capabilities, not just a single task execution.

Data and memory: cornerstones of intelligent evolution

Have you noticed that AI Agents nowadays often make the same mistakes? Behind this is actually the lack of memory mechanism. Humans can become smarter and smarter because we are able to learn from our past experiences. Existing Agent systems, even the top models like GPT-4, often behave like a "goldfish memory" when dealing with complex tasks - every conversation seems to start from scratch. A truly intelligent Agent needs to establish a continuous learning mechanism, such as through the RAG technology to transform the results of each interaction into reusable knowledge, so as to avoid repeating the pitfalls in similar scenarios.

Vertical deep plowing: the real path to landing smart

The case of Manus teaches us an important lesson: you can't have too much. The "universal Agent" that claims to be able to do everything ends up being unable to specialize in any specific area. On the contrary, OpenEvidence in the medical field, which only focuses on medical literature analysis, can give professional advice with citations. This vertical field of deep intelligence, but more practical value than those who know a little bit of everything "oil". After all, in the real work scene, we need a specific task to do the ultimate expert, rather than a know a little bit of everything layman.

Assessment systems: redefining the criteria for "smart"

Is the existing way of evaluating AI Agents really scientific?The results of the HAL team's study were shocking: a surprising 21 out of 36 test cases showed that higher reasoning effort instead reduced accuracy! This suggests that we may need to rethink how we measure intelligence. In addition to accuracy, the interpretability of behavior, the stability of task completion, and even the ability to self-correct after making a mistake should also be considered. After all, in real-world applications, an Agent that recognizes its mistakes and corrects them is much more reliable than a "genius" that occasionally makes big mistakes.

In the end, in order to achieve true intelligence, AI agents may need to abandon their obsession with "generalization" and instead pursue depth and expertise in specific areas. Just like human experts, true intelligence is not about knowing the answers to all questions, but about knowing how to find the best solutions within one's own area of expertise, and being able to continue to improve in the process. This is perhaps the most valuable lesson Manus leaves behind for the industry as a whole.