Hi,
I am developing an example using the AI Agent Element.
On the first loop, it works fine - it passes a tool out which invokes the sub process with the relevant tool in. However, when exiting the sub process and back in to the node, despite some very clear guidance in the Prompt - the second time around no tool is passed out.
Is there anything I should look for in the setup to handle this. It does follow the AI Agent Blueprint Demo though.
The cluster I am using is 8.8 Alpha, and the element is the “non-alpha” version of the AI Agent node.
@Dai_Williams - this is a great question, and one that I think everyone is struggling with while developing agents. The first thing I would look at is the AI’s reasoning. If you look at your running or finished process in Operate, and drill into the prompts, results, and reasoning - what does the AI say about it’s decisions?
Thanks for getting back to me Nathan.
I have captured the agent output for the first and second runs of the Agent element, the second time definitely doesnt populate the toolCalls in the second run. However, it does on the “toolResults” variable. Although that might remain from the initial run.
The tools available are very narrow, only 2 - send an email and create a task to review.
Very strange!
Hi @Dai_Williams - apologies for a delayed reply. One thing I’ve had some success with is taking the prompt and related details and asking the same LLM model why the prompt isn’t working as you intended it to. After a few iterations of that, I’ve managed to develop some pretty strong prompts. However, because a core feature of AI is that it is probabilistic, there isn’t a “one size fits all” fix for this type of situation.