Enterprises are moving from a handful of AI experiments toward networks of agents that can call systems, use data, and execute business logic. The result resembles shadow IT, but the untracked asset is no longer just an application. It is an autonomous actor with permissions and a job to perform.
IDC expects more than one billion actively deployed AI agents by 2029, a forty-fold increase from current levels, while agent creation rose 119 percent in the first half of 2025. At that speed, a spreadsheet inventory becomes obsolete almost as soon as it is completed.
Salesforce’s expanded MuleSoft Agent Fabric reflects the new priority: organisations must discover their agents before they can govern them.
Automated Discovery Creates a Baseline of Truth
Agents often enter through different routes. Marketing may deploy one on Salesforce Agentforce, a development group may use Amazon Bedrock, and another team may build on Google Vertex AI. Bespoke agents and Model Context Protocol servers can emerge outside central procurement.
MuleSoft’s Agent Scanners continuously inspect major ecosystems to identify running agents. They extract metadata describing an agent’s capabilities, underlying language models, and authorised data endpoints. The information is normalised into Agent-to-Agent specifications, producing a consistent profile across vendors.
That automation reduces dependence on developers remembering to register every deployment. It also gives security and operations teams current evidence rather than documentation assembled for a launch review and forgotten afterward.
Discovery is not approval, however. A catalogue must distinguish sanctioned, experimental, dormant, and prohibited agents. It should record owners and business purposes, because technical metadata alone cannot explain why access is justified.
Visibility Connects Security With Cost Control
An agent that processes loans or customer requests needs permissions that match its task. A live catalogue lets a security team see which financial databases it reaches and whether that access has expanded. Continuous monitoring can flag changes instead of waiting for the next manual audit.
The same view reveals duplication. Separate regions may be paying for several summarisation agents that perform nearly identical work. MuleSoft’s Agent Visualizer is designed to filter an estate by function, helping operations leaders consolidate tools, reduce licensing costs, and direct investment toward genuinely new capabilities.
Reuse also becomes easier. Teams can register homegrown agents and MCP servers by URL, making proprietary tools discoverable elsewhere in the organisation. Capita sees automated cataloguing as a way to spend less time managing inventory and more time building multi-agent solutions. AT&T is using the framework across support, chat, and voice interactions.
The combination is important: governance should not merely restrict agents. A reliable catalogue can help teams find trusted components, avoid rebuilding the same function, and innovate within known boundaries.
Every Agent Needs an Owner, Policy, and Exit
CIOs should assume their initial inventory is incomplete. Automated scanning can establish a baseline, but governance must then define what every bought or built agent is required to expose: purpose, owner, model, permissions, data flows, dependencies, cost, evaluation results, and current status.
Policies should control both identity and behaviour. Agents need least-privilege access, spend limits, logging, version tracking, and escalation paths when confidence is low. High-impact actions may require human approval, while credentials should expire when the owner or use case disappears.
Lifecycle management is equally critical. An unused agent with valid access remains a liability. Retirement should revoke credentials, preserve necessary audit records, and update dependent workflows. Incident response plans must account for chains in which one agent delegates to another.
Andrew Comstock of MuleSoft argues that organisations need freedom across the multi-cloud AI landscape alongside unified visibility and control. That balance will determine whether agentic systems scale constructively.
The decisive enterprise advantage may not come from owning the smartest individual agent. It may come from knowing which agents exist, what they can do, and how they interact. A coherent network can be measured, secured, and improved. An invisible one can only surprise its operators—usually after it has already spent money, exposed data, or made a consequential decision.
