Interesting, how do you handle the observability side during training? One thing I ran into with multi-agent RL is that reward signals alone don't tell you much about why an agent is failing. Curious if you've built any tooling around that.