There is nothing basic about this.

At $dayjob we were preparing some features that are aimed at agents more than humans.

When it came to evaluating how well agents can use this, we saw all sorts of novel cheating (/collaboration) attempts that made it pretty much impossible to evaluate it.