From Experimental Agents to Mission-Critical Infrastructure
Enterprises are rapidly moving from a handful of AI assistants to fleets of specialized agents embedded in core business systems. SAP and NVIDIA frame this shift as a fundamental change in the “trust equation” for AI: agents are no longer just chat interfaces, but actors that can touch systems of record, cross application boundaries and run workflows at scale. Within SAP’s Business AI Platform, agents are expected to understand roles, processes, permissions and strict data boundaries. That makes governance and security central design requirements rather than afterthoughts. At the same time, enterprises are sourcing agents from everywhere—cloud providers, SaaS vendors and open-source frameworks—creating a looming crisis of visibility. Without a single system of record for AI agents, LLMs and Model Context Protocol (MCP) servers, IT and security teams lack a complete inventory, audit trail or consistent policy enforcement across their AI estate.
AI Agent Hub: A Vendor-Agnostic Command Center for Enterprise Agent Management
SAP’s AI Agent Hub, unveiled more broadly through Joule Studio at Sapphire 2026, is designed as a neutral control plane for enterprise agent management. Originally limited to SAP LeanIX customers, the Hub now targets every AI agent, large language model and MCP server in an organization, regardless of vendor or deployment environment. Its core proposition is AI agent governance at scale: auto-discovering agents across systems, building a centralized registry and preventing untracked deployments. The registry capability is already generally available and functions as the system of record for AI assets. By treating agents, models and MCP servers as first-class, governable objects, the Hub aims to tame agent sprawl before it becomes unmanageable. SAP executives position this as a lesson learned from the early web services era, when uncontrolled proliferation created security blind spots and operational complexity that took years to fix.
LLM Inventory Control, Risk Evaluation and Identity for Every Agent
Beyond basic discovery, AI Agent Hub focuses on structured LLM inventory control and lifecycle governance. The platform not only catalogs agents, LLMs and MCP servers, it also evaluates and verifies associated workflows, capturing risk ratings and compliance mappings for each asset. The idea is that no agent moves into production without a documented governance record. Planned capabilities for the third quarter of 2026 extend this further with identity and access control via SAP Cloud Identity Services. Each agent receives a unique identity by default, enabling fine-grained authorization, data access control and auditable activity histories. AI observability, also slated for Q3, will add session-level telemetry such as health monitoring, tool-call correctness and root-cause analysis. Combined, these features give enterprises a way to continuously monitor behavior, measure effectiveness and enforce policies across heterogeneous agents rather than relying on siloed vendor tools.
Process and Agent Mining: Turning Agent Sprawl into Optimization Data
Addressing agent sprawl is not just about control; it is also about optimization. SAP is extending its Signavio process mining capabilities to the AI layer, using “agent mining” to analyze how agents actually execute workflows. Because AI agents are inherently non-deterministic, they may drift from their intended pathways or rely on human-in-the-loop interventions more often than expected. By mapping designed versus actual execution, enterprises can identify inefficiencies, compliance risks and opportunities to refine prompts, tools or escalation rules. Telemetry from AI observability in Agent Hub feeds this analysis, helping IT and business stakeholders understand who interacts with which agents, which tools are underused and where automation breaks down. In combination with LeanIX for architecture discovery and SuccessFactors for organizational context, SAP is betting that this integrated view will become a defensible moat for its AI agent governance strategy.
NVIDIA Partnership: Hardening Specialized Agents With Runtime Security
To reinforce trust in specialized enterprise agents, SAP is deepening its collaboration with NVIDIA. SAP is embedding NVIDIA OpenShell—an open-source runtime for developing and deploying autonomous AI agents—into SAP Business AI Platform as the security foundation for all SAP AI agents, including custom ones built in Joule Studio. OpenShell provides isolated execution environments, filesystem and network-level policy enforcement, and infrastructure containment to limit damage if agent logic goes wrong. SAP engineers co-develop OpenShell with NVIDIA, focusing on runtime hardening, policy modeling, identity integration and governance hooks tailored to production environments. In SAP’s architecture, OpenShell answers whether an agent action can safely execute, while the Joule Studio runtime asks whether it should happen at all. Together with AI Agent Hub’s vendor-agnostic oversight, this creates a layered defense model that allows enterprises to scale AI deployments without sacrificing control or flexibility in choosing agents and models.