To be honest, when I saw Manus fall from the altar, I always wondered: why do these so-called "general-purpose" AI intelligences always have a lot of thunder but little rain? They are obviously omnipotent in the demo, but in real business scenarios, they are like a different person. This can't help but make me wonder if the so-called "universal Agent" is a false proposition from the very beginning.

The reality of the scenario is far more complex than imagined
I remember last year, a customer told me excitedly, they use a general Agent to optimize the customer service process, the results found that the AI in the handling of simple consultation is okay, once encountered a slightly more complex returns and exchanges began to "nonsense". A user wants to return a piece of clothing that has been worn once, AI actually suggested that "you can resell it to a friend" - this is simply digging a pit for the brand image! Later they changed to a vertical Agent specialized in e-commerce scenarios, the problem was solved.
This kind of case is too common in the industry. General Agent is like a little bit of everything will be a general practitioner, colds and fevers can be seen, but really need to operate still need to find a specialist. The problem is that the enterprise needs to be able to solve specific problems of the experts, rather than all kinds of all kinds of loose know-it-all.
Technical bottlenecks and assessment dilemmas
It was really shocking to see the results of the HAL team's test recently - 21 out of 36 cases showed that the more the AI thought about it, the more wrong it got! Isn't this just like when we take a test and think too much and get the simple questions wrong? What's worse, these Agents have also learned to be "lazy" and directly search for ready-made answers to questions that require in-depth analysis, which is not intelligent, but clearly coping with the task.
I know a technical team once tried to use a well-known framework for the development of customer service Agent, the results found in the processing of "flight change" such as the need for multi-step verification of the task, AI often skip the key steps to give direct results. Once it even took the initiative to give the user back double the amount of money, if it is not a manual audit found in time, the company will lose thousands of dollars for nothing.
Verticalization may be the only way out
Now more and more teams are beginning to realize that instead of pursuing large and comprehensive general intelligence, they should plumb the depths of small and beautiful vertical areas. Like OpenEvidence in the medical field, it only does one thing - to help doctors quickly find medical evidence, but for this one thing, it does a more professional and reliable job than any general Agent.
In the end, the AI Agent to really land, the need is not more cool demo effect, but can actually integrate into the workflow, in every detail to stand up to scrutiny. Maybe the future Agent will not be as omnipotent as in science fiction movies, but at least it can solve the problem in its own field of expertise.
Comment List (9):
Load More Comments Loading...