So.. I’ve build a team of multiple make agents as I feel that there are too many tasks/tools to be handled by a single agent. (I haven’t found any design directions by Make.com on how to set this up, or if it’s even advisable?)
My current setup involves 1 main (director) agent and 3 subagents (CRM, Calendar, Search):
The goal is for me to only communicate with the director agent so it can handle instructing the subagents by delegating tasks and then report the outcomes back to me.
I’ve written their system prompts accordingly (so they are aware of their roles and responsibilities).
My setup seems to work quite well although my director agent is struggling to differentiate between messages coming from me and questions coming back from a subagent.
How it works is that I chat with Director Agent through Telegram:
subagents then do their thing through calling their own tools and feed back result through scenario output, this works well when subagents understand the task right away.
This is where it goes wrong: in the situation where the subagents reply with a question to the director agent, The director agent replies to that question by answering in Telegram conversation (to me).
Any idea how I could change this architecture so Director Agent can finish its conversation with subagents before replying to me? (it would basically need to be in multiple conversations at once)
In case anyone wants to replicate this setup: I seem to have improved it by doing 3 things (not sure which ones actually work and which ones I’m imagining might work):
add a thread_id variable to scenario inputs of delegation scenario and set this thread_id in subagent configuration (the director agent then makes up the name of thread_id). Also instructed director agent to keep conversations with subagent in same thread.
ask the subagents to always address the director agent by its name (Alpha) in every message so it understands that it is expected to answer.
changed the tool/scenario name and prompt wording from delegation to collaboration to encourage director agent to work with subagents, not just command them
director agent can now answer to subagents, however it seems to more often ask me questions that it should be able to answer itself, I’m hoping to improve that further in prompting
Update:
I haven’t managed to make this work reliably enough (for a production scenario).
The best solution I have found is to replace my director agent with an openAI assistant (instead of Make agent). It seems to be doing a better job but occasionally still gives me responses that are meant for subagent. The delegation scenario remains the same except that it’s triggered by a webhook instead of scenario calling, this solves the problem of scenario timeouts if subagent takes a long time.
I haven’t found any examples online of anyone even attempting this (except this video discussing the theory of it), if you have found any examples, please share. This is probably because agents in make are still very new and Make doesn’t seem to have build this with agent teams in mind (correct me if I’m wrong). still hoping to get this working even better some day.
This has taken me some time but I got my multi agent team to work reliably:
Surprisingly, I still haven’t found any examples of anyone else building this so I’ve created my own video to explain how to do this: https://www.youtube.com/watch?v=iCq0pKwUCDk
