Meta's CORAL puts AI agents in charge of a production recommender serving billions, with real A/B results
omarsar0 · x · 2026-09-03
Meta published CORAL, arguably one of the most convincing agent deployments to date: an agent operates against a live production recommender serving billions of users and reports real A/B results.
Sustaining a recommender is continual optimization: content, user behavior, and upstream models keep shifting, and retrieval/ranking/serving choices must be revisited — too slow for human engineers via online experiments alone.
How CORAL works:
- Each cycle, the agent observes operating signals
- Reasons over a memory of past decisions and their measured outcomes
- Invokes tools including a numerical optimizer that keeps every change within a fixed operating budget
- Improves its policy in context
A strong demonstration of agent harnesses in production-grade recommender systems.
Related event: Meta Unveils CORAL: Agentic Loop Powering Production Recommenders(2 posts)→
More from coding & agent
- Jensen Huang explains the agent harness at G20: an exoskeleton around LLMs — rohanpaul_ai · 2026-09-03
- Solo dev ships Stalkr analytics: brand mentions, share of voice, AI-sorted feedback — marclou · 2026-09-03
- Sentry Adds Support for Both Effect v3 and v4 TypeScript Framework Versions — zeeg · 2026-09-03
- NVIDIA's Object Oriented Agents: build AI agents via simple Python subclassing — s_scardapane · 2026-09-03
- Programmer satirizes total AI outage: 'I don't remember how to think for myself' — Calrose_rice · 2026-09-03
- Agents + Revid CLI: auto-generating weekly product highlight videos sent to affiliates — tibo_maker · 2026-09-03