Ryan Carson explains how Untangle’s shift toward serving family law firms created rapid growth and forced him to develop systems for managing a large number of AI agents. He organizes work through folders, priority threads, written task lists, and detailed playbooks, showing that effective delegation still depends on strong human management. He describes Watchdog, which monitors customer accounts, errors, activity, and unresolved work to identify urgent issues. The discussion emphasizes that generating more code does not necessarily produce a better product: founders must speak with customers, exercise product judgment, and avoid shipping unnecessary features. Different agents are best suited to different jobs, with cloud agents handling broader operational tasks and local tools helping with fast, UI-focused work. The speakers also discuss automated pull-request review, risk scoring, testing, walkthroughs, and merge queues as ways to reduce bottlenecks without eliminating oversight. They compare async agent threads with Slack and describe a workflow in which Claude creates design specifications while Codex builds the technical system. Finally, they advocate evaluating engineering candidates through realistic work samples rather than relying mainly on interviews, reinforcing the broader lesson that AI should be optimized for meaningful outcomes, not raw output.