Meta cut preparation time for new data column configurations to under an hour by assigning work to specialist AI agents, the company said in an engineering publication on Oct. 19.
The work previously took three to four days manually and about a day with an earlier configuration-driven system, Meta said. The company published its findings in “The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do.”
Meta said about two-thirds of the system’s changes reach production without human edits by its measure. The company attributed that result to specialists with detailed knowledge of particular tasks, while acknowledging that the design sacrifices the ability to handle more general work.
The company examined its experience automating the addition of data columns across its recommendation systems, where coordinated software changes must stay aligned. It did not specify an evaluation period or the number of tasks used to assess its own system.
The process spans three code repositories, three programming languages and up to six processing stages, Meta said. Adding a column requires changes across configuration records, backend services and the pipelines that deliver data to analytics systems, the company said.
Meta said its initial approach gave one AI agent all the knowledge needed to perform those jobs. Beyond a certain amount of information, the agent began mixing rules from different systems and applying instructions to the wrong tasks, the company said.
A separate single-agent approach succeeded on 70% to 80% of tasks when engineers wrote custom instructions for each job, Meta said. Without that assistance, its success rate fell to 10% to 20%, the company said.
The replacement uses a central router, software that assigns and sequences tasks, with 1,300 lines of code, Meta said. It coordinates more than 20 workflows through 14 specialist agents and performs none of their domain-specific work itself, the company said.
The router follows a fixed seven-step process, with only the sixth step requiring reasoning by a large language model, Meta said. That reasoning takes place inside a separate specialist agent loaded with only the knowledge relevant to its task, the company said.
Meta said it added support for moving data pipelines by arranging existing agents in a new sequence, without writing new agent code. More than 25 knowledge files supply instructions when specialists run, allowing updates without changes to the router, the company said.
The company said it is expanding the architecture into new pipeline domains and expects growth to require additional workflow instructions and knowledge files. It is also exploring automated feedback to identify failures and propose improvements, with human review required before changes take effect, Meta said.
