I have been playing around with agents and tools using Langhchain, Claude and OpenAI. Some of the tasks I prompt are similar to:<p>`Generate an investment memo on Tesla`<p>So the agent has access to search tools and financial APIs from Yahoo. What it needs to do theoretically for this prompt is first fetch updated live data and then search for recent news perhaps and then use the LLMChain and Default LLM tool to generate an investment memo that will incorporate the findings.<p>What actually happens instead: It uses only one tool. Then the data is outdated.<p>The same idea applies to a lot of things. Did anyone figure out how to chain tools like this?
We have been working extensively on this and the answer is: it is just not good enough at 'managing' agents yet to do this reliably. We are using AI to generate tool descriptions that get fed in, but making them collaborate is very hard.