Meta’s first AI agent for its data pipelines knew everything about the job, shone in demos, and began making things up once it went into production. In an engineering post published on 19 October, three Meta engineers (Plawan Rath, Akhil Jindal, and Ashwini Verma) say the cure was to strip knowledge from the agent in charge and hand it to specialists. For any team building agents, that reversal is the most useful part of the post.
The system automates the onboarding of data columns into Meta’s recommendation systems. Keeping the pieces aligned is hard: a single column crosses three code repositories and three programming languages, and can pass through as many as six stages. The first design was one agent loaded with configuration schemas, code generation rules, routing logic, validation requirements, and patterns for more than ten data sources.
Past a certain amount of context, the engineers write, extra knowledge made results worse. The agent mixed up subsystems. It applied the field naming conventions of one part of the pipeline to another, and used a code generation pattern built for one data source when filling in a different one. The post names no threshold where this began and gives no failure rate for the first version.
The redesign is a router of about 1,300 lines that handles more than 20 workflows and passes them to 14 specialist agents. The router does no domain work. It figures out the request, collects the required fields, looks for any open warnings or mandatory reviewers, and hands the job on. Six of its seven fixed steps are deterministic. Only the handoff involves a language model, and that happens inside a specialist that loads just its own knowledge files.
New capabilities mostly mean reordering. When Meta needed to support pipeline migration, it wrote a new workflow sequence from existing agents and, according to the post, added no new agent code. More than 25 knowledge files load at run time, so an edit reaches the next session without redeploying the router.
Here is what Meta measured. Configuration prep for a new column now takes under an hour. Done manually, the same work ran three to four days, and an earlier config-driven system brought that to about a day. About two-thirds of the changes the system makes go through without major human editing. Those are the only figures Meta supplies for the new design, and the post does not say what share of changes needed no edits before.
The post also claims “notable improvements” in both efficiency and accuracy, yet no accuracy figure compares the new design with Meta’s own first agent. The closest numbers describe a different single-agent approach that Meta evaluated: 70 to 80 percent success when an engineer wrote a custom prompt for each task, and 10 to 20 percent without that help. That is a comparison against a third setup, not a measurement of the thin router.
One outside source appears in the post. A preprint from the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai found that orchestrated multi-agent systems held 65.3 percent accuracy at batches of 80 tasks where single agents fell to 16.6 percent, with token use cut by up to 65 times. Meta cites it for the authors’ explanation that isolating tasks prevents interference. It studies multi-agent design in general, so it supports the logic but does not validate Meta’s results.
The scope is narrow: one team, one internal pipeline. Meta argues the pattern applies to any production agent, but the post reports no second deployment, names no model, states no cost savings, and announces no open-source release.
The practical stake is in how agent products get sold. Many pitch a single assistant that knows an entire company. Meta’s account suggests that the more such an assistant is fed, the more likely it is to cross wires, and that a mistake in the coordinating layer spreads to everything downstream. Teams running one agent with a growing pile of instructions can run a cheap test this quarter: split off the domain that fails most, give it its own context, and compare error rates before and after, the measurement Meta left out.
Engineering at Meta, “The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do,” published 19 October 2026.
