Build a Marketing Team of Agents That Runs Itself (Claude Cowork)

Most AI agents die by week three. You use it daily, then you're editing everything it gives you, then you skip it because doing it yourself is faster. The reason is always the same. You gave one agent five jobs, and a thing with five jobs does all five at about a four out of ten.

This is the exact agent team we run our agency on. Four roles, one job each. The analyst reads your numbers, the researcher records what AI says about your category, the writer drafts from the researcher's rundown, the distributor cuts that into a week of deliverables. Each hands off to the next. Four rules make it hold: one job per agent, one example of what good looks like, put it on a schedule, and it has to tell you when it breaks. One rule sits on top of all of it. Nothing publishes or spends without a human saying yes.

We built the first agent live from an empty folder to a scheduled Slack report in about 20 minutes. This is that build, packaged.

What's in the download:

• The full Claude Cowork project, with a README and the build steps

• Four agent files: analyst, researcher, writer, distributor

• The CLAUDE.md context file every agent reads first, with the human-in-the-loop governance rules baked in

• An annotated example report, plus sample marketing numbers (and a deliberately stale set to show the freshness trap)

Built by Will Leatherman and the team at Catalyst, who run their own agency on this exact system and have built content and AI-search systems for 150+ B2B companies.