Most of my agency's production work is done by 12 Claude agents written in Node.js with the Anthropic SDK. Each agent owns one discipline: diagnosis, market research, strategy, ads, copywriting, SEO, website, landing pages, PRDs, email.
An orchestrator plans the dependency graph and runs the agents in waves: research before strategy, strategy before copy. Independent agents run in parallel.
Two design choices matter more than the agents themselves. First, the human gate: after market research the system stops, posts an executive summary to Slack, and waits up to 48 hours for my approval before spending more compute. Second, auto-review: every output is scored 0 to 10 by a reviewer agent, and anything below 7 gets retried with the review notes attached.
A Notion deal moving to 'Closed Won' triggers the whole pipeline through a Make.com webhook, and client onboarding starts itself.
It runs on roughly 9 real clients today. It is not magic: the instruction files need constant maintenance, and the QA loop catches maybe 80% of what I'd catch myself. But it turned delivery from my calendar's biggest block into a process I review instead of perform.