---
title: "The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do"
description: "Meta’s thin orchestrator is a 1,300-line deterministic router that handles 20+ workflows across a data column onboarding pipeline spanning three code repositories, three programming languages, and up to six synchronized stages. It delegates all domain reasoning to 14 specialized subagents — each with isolated context loaded from 25+ knowledge files — and follows a fixed 7-step protocol that includes reliability gate checks before any LLM invocation. This architectural separation replaced an earlier monolithic agent that suffered from context interference, achieving only 10-20% success without per-task prompt engineering. The thin orchestrator pattern uses reusable workflow templates so that new use cases, such as pipeline migration, require no new agent code — only a different sequence of existing agents. Configuration preparation dropped from three to four days to under an hour, with roughly two-thirds of changes landing without significant human edits."
---

# The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do

[Meta](https://yomu.fyi/company/meta) · By Plawan Rath, Akhil Jindal, Ashwini Verma · Oct 19, 2026

**Type:** Problem & solution

## Summary

Meta’s thin orchestrator is a 1,300-line deterministic router that handles 20+ workflows across a data column onboarding pipeline spanning three code repositories, three programming languages, and up to six synchronized stages. It delegates all domain reasoning to 14 specialized subagents — each with isolated context loaded from 25+ knowledge files — and follows a fixed 7-step protocol that includes reliability gate checks before any LLM invocation. This architectural separation replaced an earlier monolithic agent that suffered from context interference, achieving only 10-20% success without per-task prompt engineering. The thin orchestrator pattern uses reusable workflow templates so that new use cases, such as pipeline migration, require no new agent code — only a different sequence of existing agents. Configuration preparation dropped from three to four days to under an hour, with roughly two-thirds of changes landing without significant human edits.

## Context

Meta’s data column onboarding pipeline for recommendation systems, spanning three code repositories, three programming languages, and up to six stages.

## Approach / What changed

The thin orchestrator pattern: a deterministic 1,300-line orchestrator that routes work using a fixed 7-step protocol (intent classification, duplicate check, field gathering, auto-discovery, reliability gate, delegation to specialist subagent, result handling). It delegates all domain reasoning to 14 isolated subagents with runtime-loaded knowledge files, using reusable workflow templates to sequence existing agents for new use cases.

## Takeaways

- The thin orchestrator pattern separates routing and reliability checks from domain reasoning, placing deterministic decisions in the orchestrator and all LLM reasoning in isolated subagents.
- When a heavyweight orchestrator hallucinates, errors propagate downstream; a thin orchestrator limits hallucination because its routing and reliability steps rely on fixed protocols, not LLM reasoning.
- Context isolation prevents cross-contamination: a subagent for one data source never sees another’s patterns, reducing hallucination of wrong conventions.

## FAQ

### How does the thin orchestrator handle failures?

If a subagent fails, the orchestrator retries or escalates. The failure does not corrupt the orchestrator's state or other subagents' contexts, containing the issue.

### How quickly can the system adapt to new use cases?

New workflows are added by writing workflow templates that sequence existing subagents differently, without building new agent code. The system achieved production capability for pipeline migration with zero new agent code.

### What was the success rate of the earlier single-agent approach?

The single-agent approach achieved 70-80% success with per-task prompt engineering, but dropped to 10-20% without. This showed the problem was architectural, not prompt engineering.

**Tags:** [AI Agents](https://yomu.fyi/topic/ai-agents), [Architecture](https://yomu.fyi/topic/architecture), [LLMs](https://yomu.fyi/topic/llm), [Reliability](https://yomu.fyi/topic/reliability)

- Source: [Meta](https://engineering.fb.com/2026/10/19/ml-applications/thin-orchestrator-ai-agent-architecture)
- Source URL: https://engineering.fb.com/2026/10/19/ml-applications/thin-orchestrator-ai-agent-architecture
- Ingested by Yomu: 2026-10-19T18:00:14.126Z

[Read original post](https://engineering.fb.com/2026/10/19/ml-applications/thin-orchestrator-ai-agent-architecture)