Meta split a single AI agent into 14 specialists after it hallucinated
Meta · 19 Oct 2026 · By the Cofoundr team
In a post on Engineering at Meta, engineers described an internal agent for onboarding data columns in recommendation pipelines. It worked in demos but mixed up rules in production, so Meta replaced it with a router that sends work to 14 narrow subagents.
The key points
- Only one of seven workflow steps uses LLM reasoning, according to the post.
- Meta says the task fell from days by hand to under an hour, but gives no sample size or error rate.
- It is internal data plumbing, not an ad product.
What it means for agencies
The lesson transfers to agency workflows: narrow, rule-bound agents with a human check tend to behave better than one agent given everything. It is a useful pattern if you are building AI routines around ad accounts.
Source: PPC Land. This is our own summary. Read the original for the full detail, and check Meta's current documentation before changing client accounts.