From the blog
Meta Ads Automation Guardrails for AI Agents
Automating Meta Ads with agents can save a lot of time — but a badly structured action can also put a valuable ad account at risk. This guide is a practical set of guardrails for using Meta API, MCP, and CLI tools safely, restrictively, and with explicit approval where it matters.
Meta ad account blocks are real — advertisers face them frequently. And the thing is, an account with history, working campaigns, audiences, and proven creatives is precious. If it gets restricted or blocked, rebuilding that setup can take months — sometimes much longer.
And then there is automation.
Since Meta released the official MCP and CLI, it has become incredibly tempting to give Codex / Claude access and let agents do more of the routine Meta Ads work. And honestly, it is amazing when it works.
But the dangerous part is not automation itself — it’s uncontrolled execution.
Agents retry things. They probe APIs. They can accidentally send too many requests, repeat a mutation, make a change against the wrong account, or just keep going when a human would stop and investigate.
That’s the part I wanted to protect against.
So I wrote — with ChatGPT Codex, for sure — a short but very specific AGENTS.md policy/template for Meta Ads automation.
The basic idea is deliberately restrictive:
- read-only or dry-run by default
- exact human approval before writes
- small, serialized write batches
- no blind retries
- explicit isolated account profiles
- one approved wrapper around the official Meta Ads CLI
- immediate stop on rate-limit / request-volume signals
- circuit breakers when the state becomes ambiguous
It’s less “please be careful” and more “these are the exact conditions under which the agent is allowed to continue”.
I built it from my own experience working with agents and Meta Ads accounts — including accounts that actually got blocked — plus patterns and experiences I found shared by other advertisers and developers online.
One important caveat: this is a community project policy, not an official Meta recommendation. It’s intentionally conservative.
Long story short, these are the instructions I point agents to when I’m working with Meta Ads or building automations around them.
It’s written in plain English, so you can inspect exactly what the agent is being told to do. It’s also MIT licensed, so feel free to use it, fork it, modify it, or adapt it to your own setup.