Building a multi-agent system around a different bet than most frameworks. Most agents jump straight to execution when you give them a task. People don't. We filter first, check if we're even the right person, look at what came before, then start.
So I've been building agents where identity is a real filter (not a label in a prompt), context is narrowed per phase of work, and an agent can reject a task and hand it back if it isn't theirs. Nine specialized ones that push back on each other on strategy work.
Curious if anyone here has tried this kind of architecture and where you ended up.
Top comments (0)