MACS framework uses deterministic agents to hit 87.1% pass rate in e-commerce recs

Most LLM-powered shopping assistants rely on the model to both understand requests and pick products, which leads to hallucinations and broken constraints when inventory is fixed. A new paper from MIT researchers (Huang, Clay, Beygi, Sarda, Golrezaei, Saberi) introduces MACS, a Multi-Agent Commerce System that splits responsibilities: LLMs handle conversation and preference elicitation, while a deterministic merchant agent performs product retrieval, hard-constraint filtering, brand exclusion, and progressive relaxation of overly strict filters. A session-persistent preference layer tracks constraints across turns, allowing consistent handling of budget overwrites and exclusion reversals—cases where users change their mind mid-conversation.

The results show a clear reliability win. On a 140-query single-turn benchmark, MACS achieves the highest pass rate (87.1%) and perfect brand compliance (1.000). On a 10-scenario multi-turn benchmark, it reaches 72% macro Pass@5, compared to 56% for GPT+Catalog and 52% for Gemini+Catalog, with zero constraint drift. The gap is sharpest on the hardest cases: exclusion reversal (100% vs 20% and 0%) and constraint accumulation (100% vs 60% and 40%). Mean judged response quality stays competitive (0.751 vs 0.736), so gains in reliability don't come at the cost of conversational quality. The paper will be presented at the Stanford Trust & Safety Conference and has already appeared at the Stanford Market AI Conference.