> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/ai-enterprise/ai-factory/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/ai-enterprise/ai-factory/_mcp/server.

# Platform Services

> Shared gateway, AgentOps, data and action integration, AI asset catalog, delivery, identity, trust, policy, observability, and detection services.

Platform services apply shared enterprise controls around AI workloads and runtimes. They should remain usable across supported model, platform, and infrastructure choices.

Explore [Platform Services ecosystem solutions](/ai-enterprise/ai-factory/ecosystem-integrations/solution-catalog?layer=platform-services).

## Service Responsibilities

Use these responsibilities and boundaries to define Platform Services. Linked documentation illustrates specific implementation capabilities; product selection remains a deployment decision:

| Service                      | Responsibility                                                                                                                                                                 | Boundary                                                                                                                                                                                                                                                                                                                                                                                      |
| ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Gateway and API              | Service discovery, authentication, authorization, routing, quotas, guardrails, sensitive-data disclosure controls, telemetry, and policy.                                      | Does not execute models or replace authoritative identity. Refer to [Inference Hub architecture](https://docs.nvidia.com/enterprise-reference-architectures/ai-factory-inference-hub-reference-design/latest/reference-design-and-request-flow.html).                                                                                                                                         |
| AgentOps                     | Agent traces, evaluation, quality, autonomy limits, human approvals, delegated-execution oversight, fault analysis, release controls, and outcome monitoring.                  | Complements rather than replaces service and infrastructure operations. Refer to [NeMo Relay observability and governance](https://docs.nvidia.com/nemo/relay/about-nemo-relay/overview).                                                                                                                                                                                                     |
| Data integration             | Read-oriented access to governed enterprise data and knowledge through approved APIs, retrieval services, or Model Context Protocol (MCP) resources.                           | Returns bounded context and protects persistent agent memory, retrieved context, and reused knowledge throughout their lifecycle. It does not grant authority to change an external system. Source systems retain ownership and authorization authority. Refer to [NeMo Retriever document ingestion](https://docs.nvidia.com/nemo/retriever/latest/extraction/workflow-document-ingestion/). |
| Tool and action integration  | Bounded execution of consequential operations through APIs or MCP tools, with explicit identity, policy, approval, credentials, disclosure controls, and audit.                | Action authority is distinct from data access. The target system remains the final enforcement point. Refer to [OpenShell policy controls](https://docs.nvidia.com/openshell/how-it-works/policies/overview).                                                                                                                                                                                 |
| AI asset catalog and lineage | Discoverable identity, ownership, versions, relationships, evaluation state, approvals, and deployment status for models, datasets, prompts, tools, and agents.                | Catalogs lifecycle metadata and evidence. Repositories retain artifact bytes, and source systems retain data authority. Refer to [NGC artifact and model registry](https://docs.nvidia.com/ngc/gpu-cloud/ngc-private-registry-user-guide/index.html).                                                                                                                                         |
| Delivery and supply chain    | Version-controlled deployment state, immutable and verified software, model, tool, and policy artifacts, and controlled promotion and rollback.                                | Owns provenance and delivery state, but does not replace independent authorization or environment-specific change approval. Refer to [NVIDIA Secure AI Operations Guide](https://docs.nvidia.com/nvidia-secure-ai-operations-guide.pdf).                                                                                                                                                      |
| Identity and trust services  | Shared identity, bounded credentials, secrets and keys, attestation, and trust verification for users, agents, workloads, and services.                                        | Implements identity and trust services while Governance and Assurance owns identity requirements, accountability, and acceptance policy. Refer to [NVIDIA Attestation](https://docs.nvidia.com/attestation/).                                                                                                                                                                                 |
| Observability and detection  | Metrics, logs, traces, deployed-asset inventory, security configuration and drift detection, and tamper-evident audit collection.                                              | Supplies signals and detections. Operations and Resilience owns triage, remediation, response, continuity, and recovery. Refer to [AI Factory Observability Guide](https://docs.nvidia.com/enterprise-reference-architectures/observability-guide/latest/).                                                                                                                                   |
| Policy and enforcement       | Policy authoring, verification, and distribution plus independent authorization, disclosure protection, delegation limits, and mediation of tool, network, and credential use. | Enforces approved requirements outside the model prompt. It does not own business policy, accept risk, or replace authoritative systems. Refer to [OpenShell policy verification](https://docs.nvidia.com/openshell/how-it-works/policies/prover).                                                                                                                                            |

## Security Responsibility Boundaries

* **Policy verification:** Check policy consistency and required security properties before distribution and after changes. Governance and Assurance defines the required evidence. Record verification coverage, assumptions, and unresolved results; a policy check does not prove runtime enforcement.
* **Memory and context security:** Protect persistent agent memory, retrieved context, and reused knowledge during ingestion, retrieval, reuse, retention, and deletion. Preserve provenance, access boundaries, and separation between tenants. Treat retrieved content as input, not authority to change policy.
* **Disclosure protection:** Permission to read data does not grant permission to disclose it. Gateway and API and Tool and Action Integration apply approved inspection, filtering, redaction, and destination controls before information leaves its authorized boundary.
* **Autonomy and delegation:** AgentOps tracks human approvals, autonomy limits, and delegated execution. Policy and Enforcement preserves those limits across delegation; a child agent cannot acquire broader authority through delegation alone. Isolation and Runtime Services supplies enforcement and containment primitives.

[OpenShell policy verification](https://docs.nvidia.com/openshell/how-it-works/policies/prover) illustrates policy-property checking within a documented model and scope. [NeMo Guardrails](https://docs.nvidia.com/nemo/guardrails/latest/index.html) provides implementation options for inspecting inputs, retrieved context, tool calls, and outputs. Select implementations against your requirements; neither example is mandatory or sufficient for every control.

## Design Requirements

* Use stable contracts between the gateway, model backends, clients, identity, policy, and telemetry systems.
* Separate human, workload, service, agent, workspace, and tool-call identities where their authority differs.
* Issue bounded, short-lived credentials rather than copying long-lived user credentials into workloads.
* Apply policy at enforceable boundaries outside the model prompt.
* Evaluate platform and workload attestation before releasing protected keys, credentials, models, or data when the threat model requires measured trust.
* Route runtime and policy telemetry into security operations with named triage, containment, patch, and recovery owners.
* Keep artifact identity immutable across build, scan, approval, mirroring, deployment, and rollback.
* Assign stable identifiers to AI assets and link source, derived assets, evaluations, approvals, deployments, incidents, and retirement state as lineage.
* Keep data retrieval contracts separate from action contracts so read access cannot silently become execution authority.
* Separate policy authorship and approval from independent decision and enforcement points. Retain attributable evidence for every consequential decision.
* Correlate workload outcomes with platform and infrastructure signals.
* Define disconnected or restricted-network substitutes before relying on a hosted dependency.

> **Info**
>
> **Boundary:** [Governance and Assurance](/ai-enterprise/ai-factory/core-reference-design/security-and-governance) defines architecture-wide policy, risk, compliance, accepted evidence, and accountable decisions. Identity and Trust Services, Policy and Enforcement, Delivery and Supply Chain, and Observability and Detection implement or evidence those requirements. The Cloud-Native Platform supplies runtime, network, admission, and isolation primitives beneath them.

## Related Capabilities

* [Secure Agent Workspace](/ai-enterprise/ai-factory/capability-reference-designs/secure-agent-workspace) extends identity, policy, credential, network, audit, and lifecycle services around persistent agent work.
* [Inference Hub](/ai-enterprise/ai-factory/capability-reference-designs/inference-hub#relationship-to-the-core) composes core responsibilities into a shared inference service. Use the integration guide to connect it to this stack and the linked publication for detailed service architecture and operations.