Salesforce today launched a platform capable of registering, orchestrating, governing and observing every artificial intelligence (AI) agent an organization deploys, regardless of where it was built or what large language model (LLM) it is based on.
Based on the MuleSoft integration framework that Salesforce acquired in 2018, the MuleSoft Agent Fabric makes it possible to not only create workflows spanning multiple agents but also provide them access to applications that are already integrated using the MuleSoft platform, says Andrew Comstock, senior vice president and general manager for MuleSoft at Salesforce.
At the core of the MuleSoft Agent Fabric is a registry through which AI agents, including servers based on the Model Context Protocol (MCP) and Agent2Agent (A2A) protocol, can be discovered. A MuleSoft Agent Broker then provides the routing capability needed to assign tasks to AI agents based on their specific function
A MuleSoft Agent Governance framework then provides the guardrails needed to enforce security, compliance and policy controls.
Finally, a MuleSoft Agent Visualizer creates a dynamic map of the agent ecosystem to surface how agents connect, interact and perform. That level of transparency is going to prove crucial as IT teams look to govern agentic AI workflows, notes Comstock. “It provides end-to-end visibility,” he says.
Salesforce views the MuleSoft Agent Fabric as a foundational element of a larger Salesforce Agentforce initiative the company launched last year, adds Comstock. Via the MuleSoft platform, it becomes possible to integrate the AI agents developed by Salesforce with AI agents developed either by enterprise IT organizations or third-party vendors, he notes.
It’s still early days so far as deployment of AI agents in enterprise IT environments is concerned, but it’s already apparent there will soon be thousands of them. The challenge is not just building and deploying those AI agents but also providing them with access to the data they will need to surface recommendations and execute tasks. The MuleSoft Agent Fabric provides the framework for unlocking that value in a way that has already been proven to scale, says Comstock.
There will, of course, be no shortage of integration frameworks for AI agents, but incumbent providers of these platforms have an advantage in terms of the number of application connectors they have already built and deployed. AI agents without access to data, after all, are only going to be of limited value.
Each organization will naturally need to determine not only how to integrate its own AI agents but also AI agents that will be deployed by suppliers, partners and customers, so the decision about which integration framework to employ should be made sooner rather than later. Otherwise, IT teams will once again find themselves trying to embed integration frameworks into applications and AI agents after they have been deployed rather than when they were initially being developed. The challenge then becomes making sure that everyone building those AI agents is aware of the integration standards they must now adhere to if AI agents are going to interact with not just each other but also sensitive corporate data.
