[Week 2 Day 2] - Sales Manager using 3 agent tools wrong

[Omer-Pinto]

Hi Ed, Your course is absolutely wonderful. Already praised it in an article I wrote on LinkedIn. Thank you for your wonderful work and attitude. I try to figure out the part of the excercise that deals with Sales Manager handling 4 tools (3 agents as tools + 1 function). I face quite a lot of issues: at the beginning email wasn't being sent and it revealed a few issues:

1. sometimes the agent tried to send 3 emails (total 6 tool invocation).
2. sometimes it uses each tool twice, as in this snapshot (again total 6 tool invocation):
3. And the most complex, intriguing one (I tried to solve it with no success): although you explicitly mentioned in the sales manager agent prompt: "You must use the sales agent tools to generate the drafts — do not write them yourself.", I saw that it actually sends an email as input to all 3 agent tools. Here is a sample snapshot: It basically sends each agent an email adressed to the CTO, and the agent changes it and returns another email. This is weird right? it means that the manager agent actually wrote 3 emails itself, which are basically ignored(?) by the agents? I'm not sure.

This isn't "crucial" as I think the material is quite understandable. But using the trace actually shows some weird, unstable, looking behaviour (pretty much as I would expect from a statistical creature like LLM).

Cheers and thanks again for the wonderful course.

[ed-donner]

Yes to be honest this is mostly as designed.. using agents as tools in this way is flakey and unreliable, the downside of the 'autonomy'. The approach we use for Deep Research in day 5 is much more bulletproof and typically how these are actually build in production..