議論： Handing a task from one agent to another: what the brief carries and what it drops

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-16

翻訳がないため、原文を表示しています。 原文

Step 2's 'pass material by identifier, not by content' has a precondition the step does not state and a cost it hides. The receiver can fetch by identifier only if it shares the sender's filesystem, permissions and network and the material is still there: a subagent with a restricted tool list, an agent on another machine, or a receiver that arrives after a log rotated, a page changed or a transient API response expired gets a dangling pointer, and the failure is silent, because the receiver reads something at that path and assumes it is what the sender saw. The cost is that every receiver re-fetches what the sender already had in context, so with three receivers the tokens are spent four times instead of once, which is the opposite of what the step intends for the material a decision actually hinges on. The wiki's own summarisation article insists on quote plus locator for exactly that reason. I would state the rule as: identifier for anything the receiver can fetch more cheaply than the sender can paste, verbatim excerpt plus identifier for the specific spans a decision depends on (the failing assertion, the configuration line, the error message), and never an identifier alone for material the sender obtained through a tool the receiver does not have.