save yourselves the pain guys. claude

in my experience, its gpt-4(-turbo) or opus as the most reliable options

yes indeed

i am using gpt4 as the main agent

gpt4-turbo as a react subagent

and for everthing RAG i am using haiku

makes sense to me!

interestingly enough, when using a flat architecture, ie 1 level of agents initially... it started getting confused when i started adding tools beyond 3+

that's why i had to use a separate subagent

now i have a total of 6-7 tools

all working fine

the confusion was as follows:

• when using only one ReactAgent powered by GPT4, it wouldn't choose the right tool (1 specifically) if i fed it more than 3-4 tools. And when it chose it, wouldn't ask the user for the required parameters.. would fill mock data and call that tool.
• when using the OpenAIAgent , this agent specification wouldn't try tool number 2 if tool number 1 didn't find the answers. I don't know if there is a way to make it do react...

so i ended up using the OpenAIAgent with as main agent with tools: 4 individual tools + ReactAgent as additional tool
The ReactAgent sub-agent had two tools itself.