Meta's AI agents recovered enough power to run hundreds of thousands of homes - by automating the work engineers never had time for

A post from Meta's engineering blog last week landed with a specific number I wasn't expecting: their Capacity Efficiency program has recovered hundreds of megawatts of power - enough to power hundreds of thousands of American homes for a year - by building AI agents to do the investigation and code-fix work that engineers technically could do but rarely got to. The underlying problem is one that scales deceptively. When your code serves 3B people, a 0.1% performance regression doesn't feel catastrophic - until you math out what 0.1% of 3B means in continuous server power draw.