Meta's thin orchestrator: 1,300-line router coordinating 14 specialized AI agents
rseroter · x · 2026-10-20
Meta's engineering team shares its 'thin orchestrator' pattern for production AI agents. Their first attempt—a single agent loaded with all knowledge to automate data column onboarding across 3 repos, 3 languages, and up to 6 pipeline stages—worked in demos but hallucinated in production. The fix: a 1,300-line orchestrator that routes 20+ workflows to 14 specialized agents while doing zero domain work itself. Key recommendations: separate routing from reasoning, design subagents as specialists, scale via workflows rather than new agents, and load knowledge at runtime. The post details how other teams can apply the same pattern.
More from coding & agent
- Cloudflare Workers often break subdomain-as-subfolder redirects for VPN users, causing 404s — gaganghotra_ · 2026-10-20
- Codex's new suggested-next-prompt is 'scary good', devs say — charliermarsh · 2026-10-20
- Stripe Atlas adds MCP support so AI agents can incorporate your startup — jeff_weinstein · 2026-10-20
- Agent inference costs spiral as GPU supply tightens — devs compare cloud, self-hosted, decentralized options — aruthesa · 2026-10-20
- Cross-agent verification: use a second AI model to review diffs before merging — umeshonai · 2026-10-20
- What's actually worth automating in customer support with AI agents? — Informal-Dust4499 · 2026-10-20