Should I be taking multi-agent systems seriously? It seems like a zombie paradigm being pushed forward by people who haven't been paying attention to recent developments (Claude Sonnet, o3, MCP, etc.) but I could be totally misunderstanding it.
Post
Conversation
To be more specific, it seems like "put all of the context into one big prompt and then have a big RL-tuned foundation model with tool-calling abilities grind away at it" has proven to be really effective. I don't know why you'd want to do something more complicated.
But I hear the big labs have people working on multi-agent systems so I am probably missing something important.
I thought the exact same way as you until a few weeks ago - the lightbulb moment for me was realizing that "multi agent" is purely a context management hack, a way to get sub-tasks done without burning a load of your precious token allowance
![David Watson 🥑](data:image/svg+xml,<svg xmlns="http://www.w3.org/2000/svg" width="96" height="96"><rect fill-opacity="0"/></svg>)
Post your reply
This is exactly what I was thinking, although that seems to mean that it could be made irrelevant if there are any major advancements in context size or online learning
That makes sense. This is different than the historic conception of multi-agent systems where the different agents were specialized models tuned for different tasks right?
Yes - that whole "the marketing expert talks to the programming expert" thing never made sense to me, since one model can be an expert at all of the above
The context management for sub-tasks thing seems generally useful to me
generate code without worrying about your database, storage, or auth.
