1,300 lines of code route tasks to 14 AI agents to slash hallucinations at Meta
Meta engineers found that giving a single AI agent all necessary knowledge caused hallucinations, so they adopted a thin orchestrator pattern to route work to specialists.
Key points
- Meta deployed a 1,300-line orchestrator that routes workflows to 14 specialized AI agents, performing zero domain work itself.
- The previous single-agent approach suffered from context interference, where the model applied naming conventions from one subsystem to another.
- Research from Mount Sinai shows orchestrated multiagent systems achieved 65.3% accuracy on batch tasks compared to 16.6% for single agents.
Meta engineers improved the reliability of their data pipeline AI agents by reducing the amount of knowledge the central router possesses, according to a post on Engineering at Meta. The company shifted from a single agent loaded with all domain expertise to a thin orchestrator that delegates specific tasks to fourteen specialized subagents. This architectural change reduced hallucinations and increased efficiency in production environments.
The risk
The primary risk identified by Meta was context interference within a single AI agent. When the agent held knowledge for configuration schemas, code generation rules, and pipeline routing logic simultaneously, it began cross-contaminating concepts. The model would incorrectly apply field naming conventions from one subsystem to another or use code generation patterns meant for one data source when populating from a different one. This degradation occurred past a certain context threshold, where additional knowledge impaired rather than improved performance.
Who is exposed
This issue affects teams using AI agents to automate complex workflows that span multiple code repositories, programming languages, and pipeline stages. Meta’s specific use case involved onboarding data columns across recommendation systems, a process requiring coordinated changes across configuration registries, backend services, and data pipelines. Any organization relying on a single agent to manage diverse domain-specific tasks without isolation faces similar risks of logical errors and reduced accuracy in production.
Cost and compliance
Meta reported that the single-agent approach required manual prompt engineering for each task to achieve a 70-80% success rate. Without this handholding, success dropped to 10-20%. The new thin orchestrator pattern eliminates the need for custom system prompts per job. A preprint from the Windreich Department of Artificial Intelligence and Human Health at Mount Sinai noted that orchestrated multiagent systems reduced token usage up to 65-fold compared to single agents, while maintaining higher accuracy on batched tasks.
What should you do?
Teams should evaluate whether their AI agents are overloaded with disparate domain knowledge that could cause context interference. Consider isolating specific tasks into specialized subagents with limited context rather than relying on a single omniscient model. Monitor production logs for cross-contamination errors where the agent applies rules from one subsystem to another, indicating a need for architectural separation.
The background in plain language
In AI agent design, a "thin orchestrator" is a central controller that manages workflow sequencing and reliability checks without performing domain-specific reasoning. It acts as a router, directing tasks to specialized subagents that possess isolated context. This pattern prevents context interference, where a model confuses rules from different domains. Teams usually implement this when a single agent’s performance degrades due to excessive knowledge, ensuring each specialist handles only its specific scope.
Background: AI agents
AI agents are software systems that perceive their environment, make decisions, and execute actions to achieve specific goals. Unlike standard models, they maintain internal state and interact with external tools, creating unique security risks around tool execution, memory retention, and autonomous decision loops that require specialized controls.
Read the full guide: AI Agents Explained: Answers to Security and Architecture Questions
What to do and how to stay safe: Meta
- Audit current AI agent architectures to identify if a single model is handling multiple distinct domain tasks.
- Implement deterministic routing layers that separate workflow management from domain-specific reasoning and code generation.
- Monitor for hallucinations that manifest as cross-contamination of rules or naming conventions between different subsystems.
- Evaluate the success rate of automated tasks to determine if manual prompt engineering is masking architectural flaws.
Step-by-step guide: AI Governance Explained: Building Rules for Autonomous Systems
General security guidance from the Vector Tribune newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why did Meta move away from a single AI agent for data pipeline onboarding?
The single agent began hallucinating and cross-contaminating concepts because it held too much domain knowledge, leading to degraded performance in production.
What is the thin orchestrator pattern?
It is an architecture where a central router performs zero domain work, instead handling sequencing and reliability checks while delegating reasoning to specialized agents.
How did the new architecture compare to the old single-agent approach in terms of accuracy?
The single-agent approach had a 10-20% success rate without manual prompts, whereas the multiagent system aligned with research showing 65.3% accuracy on batch tasks.
