Background AI coding agents — where an agent runs in an isolated cloud VM, makes changes, and opens a draft PR while you're away — have shifted from experimental to mainstream. The post outlines a three-tier mental model: interactive foreground work, parallel background sprints, and overnight backlog drains. A task is 'cloud-ready' only when it has clear scope, named files, a setup path, a test command, and a stop condition. Key pitfalls include the review bottleneck (volume of PRs rises while review capacity stays fixed), a new class of technical debt from code nobody wrote line-by-line, and security risks from over-permissioned cloud VM environments. Practical rules: treat every agent PR as a draft from a junior contributor, read the diff before the summary, never let an agent merge its own code, and keep ambiguous or exploratory work local.
