Complexity of prompt creating unreliable agent
I was wondering if anyone has any advice on how to improve the reliability of an agent we’re testing for a customer. See context and detailed questions below:
We suspect it may be an issue with the prompt:
We suspect it may be an issue with the prompt:
- Currently the prompt has 1 main workflow split into 4 sections in the prompt (authenticating the caller, performing the workflow variation #1 or #2, and giving closing confirmations) - which is about 15-20 steps in total.
- We also need the agent to be able to be interrupted during the workflow at any point and answer questions.
- We have 9 custom tools that the agent is using including answering FAQs
- what model are you using for the LLM?
- is your agent able to reliably do what it’s supposed to with instructions in a single prompt? or do you end up having to create sub prompts / sub agents that handle the separate workflows of ordering vs reservations
- do you allow your customers to give you any set of FAQs they have? is it just text dumped into one faq tool that calls another LLM?
