Skip to content
AI.info

The Pulse

Meta’s 1,300-Line Router Made Its AI Agents More Reliable, It Says

A 1,300-line Meta orchestrator routes more than 20 data workflows to 14 specialized AI agents. The system cuts new-column configuration work from days to under an hour, the company reports.

Meta’s 1,300-Line Router Made Its AI Agents More Reliable, It Says

Meta’s 1,300-line AI orchestrator performs no domain work at all: it routes requests among 14 specialized agents and enforces reliability checks. The company describes the system in a post published October 19, 2026, arguing that a smaller coordinating role helped its data-pipeline agents avoid mistakes that plagued an earlier, all-knowing agent.

Source engineering.fb.com

One agent had too much to remember

The system handles data-column onboarding for Meta’s recommendation systems, a process spanning three code repositories, three programming languages and as many as six pipeline stages. Python configurations drive C++ code generation, which feeds data-processing services and analytics systems. A single column can require coordinated changes across configuration registries, backend services and data pipelines.

Meta’s first design loaded one agent with configuration schemas, code-generation rules, routing logic, validation requirements and patterns for more than 10 data sources. In production, the company says, the agent mixed up concepts between subsystems—for example, applying one area’s field-naming conventions or code-generation patterns to another. An alternative that depended on custom prompts reached a 70–80% success rate, but fell to 10–20% without that manual prompting, Meta says.

Fourteen specialists, one fixed sequence

In the replacement design, the orchestrator classifies intent, collects required information, runs checks and hands work to a specialist. It does not reason about code, generate configurations or touch the codebase, according to Meta’s engineering account. Fourteen subagents handle defined areas such as schema management, data population and validation, with each agent’s context limited to its own domain.

The orchestrator follows a seven-step protocol, from reading an agent’s instructions and checking for duplicates to delegating the task and reporting the result. The first five steps and the final step are deterministic; the language model’s reasoning takes place during delegation, inside the selected specialist. For new use cases, workflow templates can arrange the same agents in a different order rather than requiring another agent to be built.

Hours instead of days, by Meta’s account

Preparing a new column took three to four days by hand and about a day with an earlier configuration-based system, Meta says. With the thin orchestrator, the task takes under an hour. Roughly two-thirds of the system’s changes land without significant human edits, the company also says.

Those figures describe Meta’s own system; the post does not provide an independent benchmark or detail the measurement process behind the reported results. The account also does not compare the design against other production agent systems. Its evidence is a company case study of a particular workflow, not a general demonstration that thinner orchestration will improve every agent deployment.

Complexity moves to the specialists

Meta’s proposal does not eliminate complexity so much as divide it: the central layer handles routing and checks, while specialist agents carry the domain knowledge. Knowledge files load when an agent runs and can be updated without changing the orchestrator; Meta maintains more than 25 such files. If a specialist fails, the orchestrator can retry or escalate rather than letting the error spread through other agents’ contexts.

That separation gives teams distinct places to modify workflow order, update domain knowledge or improve a specialist. Meta plans to extend the architecture to more data columns and pipeline domains, and to explore automated feedback loops with human review before changes take effect. The reported system currently routes more than 20 workflows through 14 agents.

Sources

  1. Original source · Blog post Engineering at Meta engineering.fb.com

Companies in this story

Related stories

Explore

More articles