Salesforce Expands xLAM Portfolio, Enhances AI Agent Capabilities

Salesforce is expanding its family of Large Action Models (xLAM), introducing new variants optimized for on-device deployments, limited GPU resources, and industrial use cases. The development seeks to improve accessibility, adaptability, and performance of AI agents in enterprise settings.

“xLAM now supports multi-turn, natural conversations, enabling more complex, real-world agentic tasks,” writes in a blog post Sukhandeep Nahal, Product Marketing Manager, Salesforce AI Research. “We have also expanded the model portfolio to increase accessibility and deployment flexibility across diverse enterprise environments.”

The expanded lineup follows Salesforce’s initial launch of the xLAM family in September 2024. Unlike conventional large language models (LLMs), xLAMs are designed specifically for action planning, reasoning, and tool use, offering faster inference speeds, reduced compute requirements, and lower operational costs. The model family has been engineered to power AI agents capable of understanding and executing real-world tasks through iterative, multi-step interactions known as multi-turn conversations.

Multi-turn tool calling enables agents to gather contextual information through follow-up questions before executing commands, addressing a key requirement in enterprise workflows. For example, a simple request such as canceling an order may prompt the agent to inquire about the order number or the need for a refund, allowing the system to complete the task accurately and autonomously.

The upgrade aligns with a broader trend among AI vendors working to operationalize AI agents in enterprise environments. Generative models often suffer from inconsistent performance and security vulnerabilities, particularly in critical workflows. “LLM intelligence is jagged,” says Silvio Savarese, Chief Scientist, Salesforce AI Research, to Cybersecurity Dive. “Its performance is extremely unstable in terms of consistency. This is something that we do not really want deployed in the enterprise.”

To improve reliability, Salesforce has introduced internal evaluation frameworks and performance benchmarks. These tools are designed to assess agent robustness and minimize the variability often observed in generative model outputs. Additionally, the company recently unveiled SFR-Guard, a suite of models that provide enhanced protection against prompt injection attacks and improve toxicity detection — features increasingly prioritized by enterprise IT leaders.

Security remains a core concern among enterprises considering the adoption of agentic AI. A Cloudera report from April 2025 identified data privacy and security as top priorities for IT leaders evaluating AI tools. This concern is reinforced by Gartner’s projection that by 2027, 25% of all enterprise security breaches will involve misuse of AI agents.

“We have got to find a way to build those champion agents that are highly capable and highly consistent,” says Shelby Heinecke, Senior AI Research Manager, Salesforce.

Salesforce’s expanded xLAM portfolio positions the company to offer enterprise-ready agentic solutions that address current limitations in LLM deployment. By focusing on smaller model architectures, task-specific design, and advanced interaction support, Salesforce aims to enhance operational efficiency and reduce deployment friction across diverse industry verticals.