Salesforce Releases SFR-AutoR&D Agents That Improve AI Models

Salesforce AI Research published SFR-AutoR&D, a set of agents that run research loops end-to-end to improve codebases and train AI models. The system is now being used to train and improve production embedding models at Salesforce.

How the agents work

SFR-AutoR&D agents receive a goal, a target system, and a compute budget. They formulate hypotheses, implement changes, and run controlled experiments against baselines. Improvements are retained only when supported by evidence and proven accurate and safe. The system completed one full cycle on an embedding model in roughly 80 hours.

The agents have already delivered measurable gains on SFR-Embedding V3, the model used for retrieval in Data Cloud. When pointed at the open-source EmbeddingGemma 2 model, the system produced SFR-Embedding v4 with better quality and latency. Earlier work improved public libraries including pandas, hnswlib, and USearch.

Impact on retrieval performance

Evaluated on 10 public retrieval tasks in MTEB, the system raised the nDCG@10 score by 6.13 points. A follow-up nine-hour session increased throughput by 36 percent and reduced latency by 17 percent with almost no loss in quality.

Embedding models determine which records and documents Agentforce agents retrieve from customer data in Data Cloud. Retrieval quality therefore shapes every subsequent response generated by those agents.

What it means for revenue teams

Revenue teams that rely on Agentforce or Data Cloud will see embedding models that surface more relevant customer records during agent interactions. Faster retrieval with maintained accuracy can shorten the time agents spend locating case history or product information. Teams should monitor changes in retrieval quality metrics as Salesforce applies these agents to additional models.