Three clients later, my agentic workflow is mostly a queue manager

I run solo, currently four retainer clients (two SaaS, one fintech, one agency overflow). Last quarter I tried to go full agentic: Claude Code in long-running sessions, parallel worktrees, one agent per client. The pitch was 4x throughput. Reality was I became a queue manager. Each agent finished in 20-40 minutes and dumped a PR I had to actually read. By Thursday I had 11 open PRs across three repos and no idea which ones were load-bearing. What actually worked: I cut back to one agent at a time, but pushed harder on scope. I write a tighter brief now (acceptance criteria, files to touch, what not to refactor) and let it run for an hour. Throughput is maybe 1.6x, not 4x, but review quality is back. The honest take for contractors: parallelism breaks because you are the bottleneck on context-switching between client codebases, not on tokens per minute. The agents don't carry your taste between repos. You still do.