This lesson is being recorded
Every video of this course is being uploaded to my YouTube channel. Subscribe so you do not miss it.
Bruno Ashford @brunoashfordSubscribeModule 14 · Lesson 14.2 · 14 min
A Real Case Study
See a real agentic system in daily operation, with the before, the after and the decisions that made it work.
Use ← → to move. The marker highlights text and the pencil draws on the slide; navigation locks while you annotate. Undo removes the last mark, Esc exits.
1The before
Every client report was assembled by hand from ad platforms, analytics and the CRM. Proposals started from a blank page. Follow-ups depended on someone remembering. Knowledge about each client lived in people's heads and in scattered chat threads, so when someone was off, work slowed down. I was approving almost everything personally.
2The org chart
Today the agency has specialist agents, each with its own skills: strategy and company metrics, sales proposals, finance and contracts, client messaging, design, copywriting, video, SEO, paid ads, CRM automation, data and reporting, and web development. An orchestrator agent coordinates client work and hands tasks to the right specialist. Each agent is a subagent definition with a clear description, and each owns a set of Claude Code skills written from our real SOPs.
3The router and the client files
One CLAUDE.md acts as the router: the universal rules (language, tone, formatting, what never to do), the list of skills with a one-line trigger each, and pointers to detailed rule files. Every client has its own folder with a client file describing the business, contacts, brand and goals. Before any agent touches client work, it reads that file. That single habit removed most of the "the AI got it wrong" moments.
Before doing any task for a client, identify the client from context, read their client file, and save every output inside that client's folder. If the client has no client file, stop and create one first.4Routines and the approval gate
Scheduled routines prepare a daily brief, sync meeting notes, check campaign performance and prepare reports while the team sleeps. But nothing is ever sent to a client automatically. Every message, report or proposal stops at a human approval step: the agent shows the exact text, the destination and the attachments, and a person approves, edits or rejects it. That gate is the reason clients trust the output and the team trusts the system.
5The after, and what I learned
The work that used to take the most human hours, reporting, first drafts, research, follow-ups and admin, is now prepared by agents and reviewed by people. The human team spends its time on strategy, relationships and quality. The lessons: start with the most repetitive task, write rules down every time something goes wrong, keep the router short, never remove the approval gate, and treat the system as a product you improve every week, not a project you finish.
Everything in this course comes from something I run every day. Alongside Bruno Ashford, I run a marketing agency called MediaGrowth. We serve clients across very different industries: construction, health and wellness, professional services, local service businesses in the US and in Brazil.
A few years ago, the agency ran the way most small agencies do: great people buried in reports, follow-ups, briefs and admin, with the owner as the bottleneck for every decision. Today it runs with a team of about fifteen specialist AI agents working alongside the human team. Let me show you how it is built, what changed, and what I would do differently.
Key points
- Specialist agents with clear descriptions beat one agent that does everything.
- A short router CLAUDE.md plus a file per client removes most context errors.
- Routines prepare work overnight, humans approve everything that leaves.
- Every mistake becomes a written rule, so the system improves every week.
Avoid
- Copying the size of my system on day one instead of starting with one or two agents.
- Skipping the per-client context file and wondering why outputs are generic.
- Letting any agent send to a client without human approval.
- Building agents before writing down the SOP they are supposed to follow.
A client meeting ends. The transcript is synced automatically, the orchestrator extracts the action items and routes them: a landing page change to the development agent, new ad copy to the copywriting agent, a budget question to the ads agent. Each agent drafts its part inside that client's folder using the client file for context. By the end of the day, a team member reviews each draft, approves the client update message, and the tasks are logged on the client's calendar. The owner was not needed for any of it.