Pretty expected. Agents need to
- read the chat history, call the llm to write a response or tool to use
- call the selected tools if any
- call the llm to interpret the tool response, and either write a final response or another tool call
- if theres another tool call, repeat as needed
You could write your own agent if can think of a faster way to process. I did have an example of a from-scratch agent using workflows that would be easy to customize