Essay: when AI agents compete, a 'local loyalty paradox' hollows out markets
TMTPost published a commissioned commentary by 观潮 arguing that when every party deploys AI agents to bid, apply for jobs and negotiate, competition's three functions — discovering prices, selecting the better party and keeping entry channels open — can be emptied out without anyone exceeding their mandate or acting in bad faith, a pattern he calls the "local loyalty paradox." The essay draws on the FTC complaint's allegations about Amazon's "Project Nessie" pricing algorithm (allegations, not court findings), simulations in which Q-learning agents learned to sustain prices above competitive levels, a Freelancer.com study finding that after an AI cover-letter tool launched, cover-letter fit's correlation with callbacks fell about 51% and with hiring about 79%, and a small Anthropic experiment in which 69 employees traded via Claude agents — stronger-model agents closed about two more deals and sold the same item $3.64 higher on average. The author notes these are mostly simulation or small-sample results and cannot be read as measurements of the real economy.
The author argues that even when every AI agent stays strictly loyal to its own principal and breaks no rule, the combined effect can still raise costs for everyone; the cited evidence is largely simulation or small-sample work that the author says cannot be treated as a measurement of the real economy.
（本文作者为 观潮，钛媒体经授权发布）
文 | 观潮
序
凌晨两点十七分。人早已入眠。
两家资金充足的机构，正在争夺同一项稀缺资产。白天，人各自设定了目标和上限；夜里，Agent替他们盯着报价，计算对手的反应，在授权范围内调整出价。
甲方交代：“别错过，价格合适就继续跟。”
#Anthropic#Claude#Freelancer.com#Amazon#Federal Trade Commission