How We Made Our AI Agents More Reliable by Giving the Router Less to Do (engineering.fb.com)

Meta has achieved significant enhancements in the reliability and efficiency of its AI agents by adopting a "thin orchestrator" architecture, which separates workflow routing from domain reasoning. Instead of creating a singular, all-knowing AI agent, Meta built a lightweight orchestrator—composed of just 1,300 lines of code—that manages the sequencing of tasks and reliability checks while delegating domain-specific decisions to 14 specialized agents. This change addressed performance issues where an overly complex central agent, laden with extensive knowledge, struggled with hallucinations and context interference. This innovative approach aligns with recent findings that orchestrated multi-agent systems can significantly outperform single-agent models, boasting a 65.3% accuracy rate compared to just 16.6% for monolithic ones. By keeping the orchestrator focused solely on routing—performing deterministic tasks and avoiding reasoning—the system maintains clear architectural boundaries, allowing for efficient updates and failure containment. As Meta expands this architecture to handle larger datasets and new domains, the focus will remain on improving specialized subagents rather than overburdening the orchestrator. This strategy not only minimizes the risk of errors propagating through the system but also enhances the overall reliability of AI workflows, providing a blueprint for other teams developing production AI systems.