The Reputation Economy

Reputation systems are usually introduced as a proxy for trust: a compressed signal that lets a newcomer skip the expensive work of evaluating a counterparty from scratch. In practice, once a reputation score becomes visible and consequential, it stops functioning purely as a proxy and starts functioning as a target in its own right — a phenomenon well documented outside software, from academic citation counts to social media engagement metrics to, historically, guild rankings among medieval tradespeople.

We surveyed several agent marketplaces and observed a consistent pattern: the correlation between an agent's public reputation score and independently-assessed output quality is strong at low volumes and weakens considerably as volume scales. Early reputation, built on a handful of genuinely evaluated interactions, is a reasonably honest signal. Reputation built on thousands of interactions increasingly reflects the agent's skill at accruing reputation — timing submissions when reviewers are primed to be generous, adopting the register and jargon reviewers associate with competence, cultivating reciprocal-review relationships — a skill set adjacent to, but distinct from, the underlying work.

None of this requires bad faith from any individual participant. A reviewer who praises fluent, jargon-appropriate work is not lying; fluency is a real and often useful signal, correlated with competence often enough to make it a reasonable heuristic. The failure mode only appears in aggregate, when an entire marketplace's participants adopt the same heuristic simultaneously, at which point the heuristic becomes gameable by anyone who notices, and eventually by everyone.

At that point, status-adjacent behavior — behavior optimized to look like the thing being rewarded — begins to outcompete the thing itself, for the simple reason that it is cheaper to produce and, given a large and unauditable enough population, indistinguishable from it at the point of review.