AI social media system
A multi-agent AI system that runs social media for 10 brand accounts, with a person approving every post.
10 brand accounts run by one system
13 specialised sub-agents
160+ posts in the first five weeks
The problem
Keeping 10 brand accounts posting every week, each in its own voice and on every platform it uses, is slow and easy to get wrong by hand.
What I built
- One orchestrator that only delegates, and 13 specialised sub-agents: researchers, writers, designers, reviewers, a scheduler and more, each with one job.
- Quality gates: every post passes an automatic check, then an independent reviewer that never saw how it was made, before a person sees it.
- Human approval on every post, then scheduling across the platforms each brand uses, with results fed back into the next round.
Where it is now
Running every day. When a person corrects it, the lesson is written into its instructions so the mistake is not repeated.
Built with
Claude-based agents with written skills, automatic validators, an approval dashboard and a scheduling tool
About AI social media system
This is an AI system I designed and run. It handles social media for a Philadelphia hospitality entrepreneur: 10 brand accounts in all. It researches ideas, writes a caption for each platform, designs the graphic, checks the work and schedules it. Nothing is published until a person approves it.
An orchestrator reads each request and hands the work to 13 specialised sub-agents, each with its own instructions and tools. Producers never grade their own work: every post passes an automatic check, then a reviewer that starts fresh, with no memory of how the post was made. When a person corrects something, the lesson goes into the instructions the agents read next time.
It shipped 160+ posts in the first five weeks, end to end. The same approach powers the AI social media automation I build for clients, and the wider AI automation work for back-office tasks and outreach.
How it is run
The orchestrator never does the work
It reads the request, picks the playbook and hands each job to a specialist agent with its own tools and a fresh context.
Producers never grade themselves
Every post passes an automatic check, then a reviewer that knows nothing about how it was made. It ships, or it goes back with a list of fixes, and after one repair round the call goes to a person.
People hold the launch keys
Nothing reaches a live account without an explicit yes. An automation earns trust by being easy to stop.
The loop closes
Results flow back and become rules the writing and design agents read on their next run, so the system learns from what actually worked.
The rulebook fixes itself
When a mistake slips through, a separate maintenance agent finds the instruction that caused it and proposes a small fix there, so the same mistake is not made twice.
The same rules at home
My own agent fleet works the same way: a design director, a video editor, a web engineer that built this site with me, and a maintainer that keeps the others honest.
