As enterprises race to deploy AI agents across cloud platforms and business functions, a new problem is emerging alongside the promise of automation: visibility. Salesforce is positioning its latest expansion of MuleSoft Agent Fabric as a response to what many CIOs are beginning to call agent sprawl—the unchecked proliferation of AI agents operating beyond central oversight.
The scale of the issue is only expected to grow. According to IDC, the number of actively deployed AI agents is projected to cross one billion globally by 2029, a fortyfold increase from 2025. Many of these agents are being created by individual teams using different platforms, often without consistent governance, raising concerns around shadow AI, compliance, and operational risk.
Salesforce’s answer is to extend MuleSoft Agent Fabric with automated agent discovery, designed to give organisations a single control plane to identify, catalogue and manage AI agents and tools across a fragmented, multi-cloud environment.
From manual audits to automated discovery
At the heart of the update are new Agent Scanners, which automatically detect and register AI agents running across platforms such as Salesforce Agentforce, Amazon Bedrock, and Google Cloud Vertex AI. The scanners continuously monitor connected environments, identifying new or updated agents and capturing technical context without manual intervention.
For IT teams, this replaces what is often a slow, error-prone process of tracking agents through spreadsheets and point audits. An inventory forecasting agent built on Vertex AI, for example, can now be automatically discovered and catalogued alongside a customer support agent created in Agentforce—without engineers having to hunt through multiple cloud consoles.
Brad Ringer, Enterprise & Integration Architect at AT&T, described MuleSoft as “a massive accelerator” for long-term AI strategy, noting that a structured framework is becoming essential as AI adoption accelerates.
Governance through metadata, not guesswork
Beyond simply identifying agents, Salesforce is emphasising deep metadata extraction. The scanners capture details such as what actions an agent can perform, which large language models it relies on, and what data it is permitted to access. This information is normalised and mapped to Google Cloud’s agent-to-agent (A2A) protocol card specifications, helping organisations establish a consistent understanding of how agents reason and act.
All discovered assets are synchronised into the MuleSoft Agent Registry, a continuously updated catalogue intended to serve as a real-time source of truth for security, governance and development teams. Homegrown agents and Model Context Protocol (MCP) servers can also be registered via simple URLs, ensuring internally built tools are not excluded from oversight.
Andrew Comstock, SVP and GM at Salesforce MuleSoft, said the goal is to let organisations innovate freely across platforms while retaining the visibility and control needed to scale safely. “The most successful organisations of the next decade will be those that harness the full diversity of the multicloud AI landscape,” he noted.
Towards a governed Agentic Enterprise
The broader ambition behind the update is to support what Salesforce calls the Agentic Enterprise—an operating model where AI agents work alongside humans across functions such as inventory, logistics and customer service. Tools like MuleSoft Agent Visualizer now provide a consolidated view of an organisation’s entire AI footprint, allowing teams to filter agents by platform, capability or environment in just a few clicks.
As AI agents multiply across enterprises, Salesforce’s move reflects a growing recognition that scale without governance can quickly become a liability. Automated discovery and cataloguing may not slow the pace of agent adoption—but they could determine whether that growth remains manageable, secure and aligned with business outcomes.