Agents exhibit variable performance, require capability-based matching, and create value at resolution not runtime. These are labor characteristics, not software characteristics. But we're pricing them like SaaS. Akerlof's lemons problem explains why benchmark saturation (everyone scores 90%+) is killing quality differentiation. Holmstrom's work on moral hazard explains why outcome-based pricing (Sierra's model) will win over per-token pricing. The gig economy is probably the closest structural analog. Curious what HN thinks about where the infrastructure layer settles.