Case Study

A 4-agent system
that runs an agency.

A 4-agent Claude Code system that runs a social media manager's entire client workflow: weekly briefings, content pipelines, client reports, and research. Built for Camille Guillain.

Status Live · Client Delivered
Client Camille Guillain, Paris
Category Multi-Agent System
Built 2026
At a glance

What is AI Social Media Operating System?

The AI Social Media Operating System is a custom 4-agent Claude Code build for Camille Guillain, a Paris-based social media manager running 4 to 7 client accounts simultaneously. It handles weekly briefings, content pipelines, client reports, and ongoing research, eliminating 70-80% of the manual repetitive work while keeping every output in the original human voice.

The Problem

What was broken.

Camille Guillain runs social media for 4 to 7 clients at a time out of Paris. The work is not complicated: weekly briefings, content drafts, monthly client reports, ongoing industry research. But the volume is crushing. At 6 clients, the Monday briefing alone was a 4-hour exercise: reading every client\'s industry news, tracking competitor moves, and writing a structured summary each of them could actually use.

She had tried generic AI tools. ChatGPT could write posts, but not in the voice of a specific client. It could summarise articles, but not remember who said what last month. It could draft a client report, but not pull the actual data from her workspace. Every time she saved 20 minutes on generation, she lost 30 minutes editing. The math was not working.

The goal was not to replace her. The goal was to remove the repetitive layer underneath her work so she could focus on strategy, relationships, and the judgement calls that actually differentiate a real agency from a content mill.

The Approach

What was built.

Instead of a single "do everything" automation, I designed four specialised Claude Code agents, each with its own scope, memory, and brand context per client. The agents share a unified Obsidian vault as their memory layer, so everything one agent learns is available to the others.

Each client gets their own brief document inside the vault: brand voice, industry context, preferred tone, content themes, and a rolling log of every piece of work the system has produced for them. The agents read the brief before they do anything, which means the output is per-client consistent without Camille rewriting every time.

The system runs on schedule (the weekly briefing fires every Monday at 8am) and on-demand (she can trigger a content pipeline for any client at any time via natural language). Every output lands in the review queue first. Nothing publishes without her eyes on it.

How It Works

Architecture in plain English.

01
Weekly Briefing Agent
Fires every Monday morning. Pulls industry news and competitor moves for every client in the vault, reasons about what matters, and delivers a structured per-client briefing straight to Camille's inbox.
02
Content Pipeline Agent
Triggered on demand per client. Reads the brief, reviews the rolling content log, drafts platform-specific posts in the client's voice, and queues them for review. Handles LinkedIn, Instagram, and Twitter/X formats.
03
Client Report Agent
Fires monthly. Pulls performance data, synthesises it against each client's goals, drafts a structured monthly report, and flags anything that needs Camille's attention before it goes out.
04
Research Agent
Continuous background worker. Monitors topics per client, saves relevant findings into the client's brief, and builds up a domain knowledge base the other three agents reuse.
05
Shared memory layer
An Obsidian vault acts as the shared brain. Per-client briefs, content history, research findings, and decision logs all live in one structured place every agent can read and write.
Try It

See it in action.

Sample Camille AI interaction

This is a sample. The live system reads each client's brand brief and content history.

Stack

Built with.

Claude Code 4 Custom Agents Sub-agents Claude Sonnet 4.6 Obsidian (memory) Markdown Scheduled triggers Python
Outcomes

What changed.

70-80% of manual work automated
4-7 clients running in parallel
4h → 30m Monday briefing time saved
4 specialised agents collaborating

The most important outcome is not the time saved. It is that Camille is now able to take on more clients without losing the personalised, human quality her existing clients pay her for. The system handles the repetitive layer. She handles the judgement. That split is what makes the build worth it.

Free Consultation

Want a system
like this one?

Book a free 30-minute call. We map your situation, scope a similar build, and agree on a fixed quote before anything starts.

Book Free 30-min Call