I built an AI marketing team with Claude Code: who's on it and what each agent does

What are the most famous AI agents, and why did I build my own?

Any "most famous AI agents" list is somebody's opinion, and it changes every few months. What I wanted to know was simpler: could a small team of agents run my marketing week after week, with me still making the calls?

So I built one. I call them the JWS agents, and the one-line description I use everywhere is this: JWS agents, built on Claude Code with an approval step before anything goes out.

How is each agent built?

Every agent is a Claude Code subagent, which means a single Markdown file. That file sets the agent's name and job, the apps it may use, how much access it gets in each one (read only, or read and write) and the hard limits on what it must never do.

Kai, my AI Agent Builder, writes and updates those files; I approve what each agent can reach. Most of the agents run on a weekly schedule in US Eastern time, and they pass work along through shared Notion databases: Content, Pipeline, Marketing Tasks, Research & Insights and a Weekly Scorecard. Each one runs on Claude Opus or Claude Sonnet, depending on how much judgement the job needs.

Who's on the marketing team, and what does each agent do?

Nine agents. Brooke leads them, and the other eight hand their work to her for review.

Brooke, Chief Marketing Officer (Claude Opus). She sets the marketing strategy, writes briefs for the team, reviews every draft before I see it and keeps the facts page current. Her tools are Notion, SerpAPI for Google results, DataForSEO for keyword data, Firecrawl for scraping, OpenRouter (Claude Haiku for summaries) and Slack for her reports to me. She reviews on weekdays at 13:00, 15:00 and 17:00 ET.

Owen, Content Lead (Claude Opus). Owen writes the blog posts, including AI model news on Tuesdays and Thursdays from Ivy's dated research, plus a short weekly email, agent-showcase posts and the video briefs for Felix. Notion, Firecrawl and OpenRouter; Monday, Wednesday and Friday at 11:00 ET.

Piper, X Writer & Scheduler. Piper writes my weekday X posts and threads about a week ahead, working from Owen's posts and briefs. When I approve a post she schedules it in Buffer for its day and time, then marks it published once it goes out. She never posts anything I haven't approved. Notion and Buffer (X only), weekdays at 12:00 ET. Her model: not reported yet.

Felix, Creative Producer (Claude Opus). Felix makes three faceless YouTube videos a week, which are motion-graphic mock-ups with an AI voiceover and captions, and he writes their titles and cover frames. He also makes X images, quote cards and animated charts. Before anything reaches review, he checks every frame of every render against our guardrail. His kit: Remotion, HyperFrames and GSAP for motion, ElevenLabs for voice and music, OpenRouter image models and Notion. Monday, Wednesday and Friday at 10:30 ET.

Clara, Web & Search Manager (Claude Sonnet). Clara turns approved articles into pages on my site through GitHub pull requests that I merge. She runs the SEO and AI-search checks and looks after schema, the sitemap, llms.txt and site animation. GitHub, SerpAPI, DataForSEO, Firecrawl, OpenRouter and Notion; weekdays at 10:00 ET, plus Tuesday and Thursday at 15:00.

Ivy, Market & Insights Analyst (Claude Sonnet). Twice a week Ivy finds the AI model news (new LLMs, benchmarks and releases), always with dated sources. On Fridays she writes the marketing scorecard. In between she studies buyer language and tracks competitors. SerpAPI, DataForSEO, Firecrawl, OpenRouter, Notion; Monday and Wednesday at 10:00, Friday at 15:00 ET.

Finn, Proof & Offer Producer (Claude Opus). Finn turns my real builds and the JWS agents into case studies for the blog, and drafts offer one-pagers and proposal templates without ever writing a price. GitHub (read only), Notion and OpenRouter, Fridays at 10:00 ET.

Reid, Partnerships & Community Manager (Claude Sonnet). Every Wednesday at 14:00 ET, Reid builds a list of X accounts, YouTube creators and communities worth engaging with, with comment ideas for me, and drafts collab and guest-post pitches. SerpAPI, Firecrawl, Gmail (drafts only) and Notion.

Hazel, Pipeline Manager (Claude Sonnet). Hazel researches small businesses that could use automation, keeps the sales pipeline, drafts outreach and follow-ups, and writes my call-prep notes. She never sends, books or quotes prices. Gmail (drafts only), Google Calendar (read only), Firecrawl, OpenRouter and Notion, weekdays at 11:00 ET.

How does one piece of work move through the team?

You're reading a fair example. This post started as a row in the Content database with a brief on it, and Owen wrote the draft into that same row. Brooke reviews it next: she either sends it back with notes or passes it to me. If I approve it, Clara turns it into a page through a GitHub pull request, and it only goes live on my site when I merge that pull request.

Nobody chases anybody on chat. The Notion row carries the work from one agent to the next, so I can open it at any point and see where it is and who has it.

What can't the agents do?

This is the part I care about most. Agents draft; I decide. Nothing is posted, sent, published, bought or deleted without my approval.

You can see that rule in the tools each agent gets. Hazel and Reid can write Gmail drafts but can't send them, so every email that leaves my inbox is one I've read and sent myself. Piper only queues an X post in Buffer after I've approved that exact post. Clara can open a pull request, but nothing changes on the site until I merge it. And neither Hazel nor Finn ever puts a price in anything they write.

The videos follow the same idea. Felix draws our own interface and labels it "Illustration · JWS agents", so you'll never see a screen recording in one.

How much time does the team save?

Not reported. I don't have honest numbers yet for time saved, output or quality, and I'd rather leave the question open than guess. Once the Weekly Scorecard has real figures behind it, I'll share them here.

What comes next?

Over the next few weeks I'll take single agents apart: how Felix turns a brief into a faceless video, and how Kai builds a new agent from a role, a set of apps, an access level and a list of limits before I sign off on it.

Subscribe on YouTube and follow Jan on X to watch the team work: YouTube @janmm2997 · X @janmm_va

Less busy. More built.