Running 15 retainers alone: the architecture decisions that made it possible

This is not a story about working harder. It’s a story about treating agency operations as an engineering problem.

Decision one: hub-and-spoke, not sequential. Most agentic pipelines I see are sequential, task A feeds task B feeds task C. This is fragile. One failure blocks the whole chain. I run a supervisor-router model: one orchestrator that knows client context, dispatches to specialized agents, and handles failures independently. If the content agent fails, the SEO research agent keeps running.

Decision three: humans stay in the loop on decisions, not execution. I review output samples. I approve strategy shifts. I handle client relationships. I do not write briefs, pull SERP data, format reports, or draft content. The system does that. My cognitive load is proportional to the number of decisions per week, not the number of deliverables.

The result: 15 retainers, one operator, around 25 hours of human work per week. The rest is automated. The agency is the proof of concept.