The community member who posted the original question is wondering if agents are necessary, as they believe they can achieve the same results using prompt engineering and API calls. The comments suggest that agents automate decision-making and iterate until a task is completed, essentially encapsulating a series of API calls and pipelines. Some community members explain that agents can be useful for tasks like the Retrieval Augmented Generation (RAG) model, where they can handle things like querying multiple indexes, inferring metadata filters, and asking sub-questions to answer the original query. However, other community members express concern about losing control over their pipelines when using agents, as the logic is largely encapsulated. Overall, the discussion centers around the utility and trade-offs of using agents versus manually building custom pipelines.
Are agents really necessary? Because, from what I've seen, I can do anything with just prompt engineering and normal api calls... or by doing it manually...
example; a RAG agent.
I can go to the vector store and use the right api calls to get the embeddings, nodes, AI api calls etc.
Agents seems useless and just a beautiful word to custom pipelines... can someone tell me the utility of them?
In the case of RAG, you might have many indexes, or want to automatically infer metadata filters or write sql queries, or you need to ask many sub-questions to answer the original query. This is all agentic stuff.
The general idea being you can hardcode every code-path for every query
I see your point. So, underneath the agents, it's just basic iteration and references to functions (or "tools") that will be called after the iterations, correct?
I fell like I lose control over my pipelines by using agents because almost all the logic is encapsulated and done. When it's the case of building my own agent, I just write simple api calls and pipelines to do something until it didn't reach the result I wanted
