10-12-2024

As I have reviewed some AI agent frameworks or multi-agent frameworks (CrewAI, AutoGen, AutoGPT,BabyAGI, Swarm, etc.)

, the most impressive feeling is that the programmable, agent-based way to manipulate LLMs could be more applicable to vertical fields.And the current framework seems to be limited in ReAct principle presented by Shunyu Yao.

That's to say that AI agents may do well in the "to B" domain, and not commonly like previous Web2 products like Facebook, YouTube, and so on.

For vertical markets of AI agents, we would need promising impact to promote productivity. But what is important to build a vertical agent is that we must have domain knowledge and domain professionals or specialists. For some common situations like reserving an airplane ticket or planning a trip, it would be useless or a kind of overthinking hallucination of the company.

I don't know whether you know RPA, which is also known as Robotic Process Automation, but to me, the agent is very similar to this. After all, AI agent behaviors could also be completely implemented by humans using tools.

So, it presents one problem: "Only if you know how to program can you use RPA." Therefore, you cannot rely on customers to create AI agents. So,the low-code agent builder platform is limited.And RPA indeed only has effect in vertical fields.

If user could use one-prompt to create an agent system to solve task ,and I guess AGI may already thereby.

It is "to B" anyway.That would be the destiny of task-sloving agent ?

edited 11-12-2024