Indeed, I think your framing of 'invariants' is a useful pattern for agents to have a reference point for them to converge towards. I've noticed that whenever I use agents, they are most effective when the goal is clearly defined and verifiable. This tends to to naturally steer them in the correct direction.