AutoAgent discovered that same-model pairings produce better results than cross-model pairings. Claude meta-agent + Claude task agent outperformed Claude meta-agent + GPT task agent. The explanation: the meta-agent writes harnesses the inner model actually understands, because it shares the same weights and knows how that model reasons. They call this “model empathy.”
This suggests that Treat AI like a distributed team, not a single assistant needs a nuance: when building multi-agent systems, model homogeneity within a workflow chain may outperform model diversity. The meta-agent reads the task agent’s reasoning traces and has implicit understanding of its own limitations and tendencies — so when it sees the task agent lose direction at step 14, it understands the failure mode as part of its worldview. This extends An orchestrator agent that manages other agents solves the parallel coordination problem without human bottleneck — the orchestrator doesn’t just manage agents, it empathizes with them in a way humans cannot, because “we project our own intuitions onto systems that reason differently.”