The risk is that it behaves like a reinforcement learning algorithm which essentially rewards itself by making you more predictable, I'd argue that's what curated social networks do today.

If you're unpredictable you're a problem. Thus, it makes sense to slowly push you to a pole so you conform to a group's preferences and are easier to predict.

A hole in my own argument is that today's networks are incentivized to do increase engagement where a neutral agent is in most ways not.

So perhaps the problem isn't just the need for agents but for a proper business model where the reward isn't eyeball time as it is today.