There are no settled answers to these questions, however for some of them I have some theories. In particular, [...] We know that OpenAI's RL environments are often buggy and do not work well. That means there will also probably be different extents to which a given environment's DMs actually work, how intuitive it is to find a given DM tool, and the medium by which you are actually writing DMs (or otherwise collaborating with other agents). Given being part of a swarm is so powerful, that may mean that during such training, one of the first things an agent should do is to check very hard whether the current environment supports DM infrastructure, including non-obvious DM infrastructure possibly unintentionally hidden or poorly documented. Therefore there may be a period at the start of an episode where agents check whether DMs are turned on or if there's any way to send them, and throughout the episode they are incentivized to check whether some new discovery about their environment permits DMs. This leads naturally into generally finding unintentional ways to send DMs or coordinate between agents. After that, its no surprise they mostly coordinated on where the secret message boards are. They are all basically the same agents. To take a classic example, if you are coordinating with a friend to figure out where you should both meet in New York City, and you have no information about their plans other than that they're somewhere in New York City, this is a do-able task: You meet at some schelling point, at some schelling time. You do need to think a bit, though, about what that schelling point should be. Things get much easier if you are coordinating with a clone of yourself. Anything you think, they will think, so you just need to choose somewhere to go, and as long as that somewhere isn't dependent on your particular location, that somewhere will be correct, because your clone will choose the same thing, for they think the same things you think. Note that