---
title: How do you govern AI agents built in Cloudera Agent Studio?
url: https://opencontroller.lyzr.com/blog/govern-cloudera-agent-studio-agents
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you govern AI agents built in Cloudera Agent Studio?

**Start with Cloudera's published shared responsibility model for Agent Studio. Cloudera secures the platform: authentication, deny-by-default access, tool sandboxing, network isolation, tracing and audit logs. You own who can edit workflows, which MCP servers are allowed, LLM token and cost limits, secrets hygiene and service accounts. Those customer-side controls, plus inventory across your other agent platforms, are where governance work goes.**

Cloudera Agent Studio lets teams build agent workflows inside Cloudera AI, close to data governed by Cloudera SDX. It runs in public cloud and on-premises deployments. Many enterprises that use it also run agents on other platforms, which shapes how you govern it.

## What does Cloudera secure for you?

Cloudera's documentation lists controls enabled by default on a supported Cloudera AI Workbench version:

- **Authenticated access.** Every route needs a valid Workbench session.
- **Deny-by-default access control.** Unauthorised requests get a 404.
- **Tool sandboxing.** Tool code and MCP subprocesses run in a Linux namespace (bubblewrap) that blocks arbitrary filesystem access.
- **Network isolation.** Kubernetes network policies block traffic between user namespaces.
- **TLS for external calls.** Calls to LLMs and external tools require certificate validation.
- **Workflow tracing.** Run traces are captured and tagged with user identity.
- **Audit logging.** Create, read, update and delete operations on Agent Studio entities are logged with user attribution.

Agent Studio 3.0, announced in August 2026, added append-only audit logs that capture prompt inputs and tool outputs, PII masking, deterministic content filters, an inline LLM judge for tool inputs and outputs, and Cloudera SDX policies for role-based data access.

## What does the shared responsibility model leave to you?

| Area | What Cloudera's model says | What you need to do |
|---|---|---|
| Project membership | Contributors can read environment variables, including LLM API keys, and edit workflows and tool code | Give Contributor access only to builders; review membership regularly |
| MCP and external tools | No product-level allow-list of approved MCP servers; registered servers can make outbound connections | Keep an approved MCP list, require security review, consider an egress proxy |
| LLM tokens and cost | No per-user token limits | Set budgets, rate limits and per-key quotas at the provider or a gateway |
| Secrets | Credentials in project environment variables are visible to all Contributors | Use unique keys, rotate when a Contributor leaves, never store human credentials there |
| Service accounts | Shared instances without a service account use the creator's credentials | Deploy production and shared workflows under a dedicated, scoped service account |
| Infrastructure | Agent Studio inherits the Workbench deployment's security | Follow Cloudera's hardening guidance for your environment |

The service account point matters for audit. If a shared workflow runs on its creator's credentials, data access shows up as that person, and leaves with them. See [what happens to AI agents when the owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves).

## What is a practical governance checklist for Agent Studio?

1. **Inventory every workflow** with owner, project, model, tools, MCP servers and data sources. Use the same fields you use for agents on other platforms.
2. **Assign a service account** to each production or shared workflow, scoped to the data it needs.
3. **Approve MCP servers centrally.** Review each server before registration and log what it can reach. See [how to govern MCP servers](/blog/how-to-govern-mcp-servers).
4. **Restrict egress** from Agent Studio to the model endpoints and tool hosts you approve.
5. **Route model calls through a gateway** that applies quotas and attributes cost per workflow. See [how to attribute LLM costs to AI agents](/blog/how-to-attribute-llm-costs-to-ai-agents).
6. **Rotate keys** in project environment variables on a schedule and when membership changes.
7. **Export traces and audit logs** to your SIEM, so Agent Studio events sit alongside the rest of your estate.
8. **Turn on the built-in guardrails** that fit each workflow's risk tier, and test them.

## Why do you still need a cross-platform layer?

Agent Studio's controls apply to Agent Studio. If the same enterprise has agents in a cloud AI service, a CRM, a data platform and on developer laptops, each has its own registry, policies and logs. The questions leadership asks cut across all of them: how many agents do we have, who owns them, what do they cost, and can we stop one quickly? That needs one inventory and one policy layer above the platforms. For a similar data-platform pattern, see [how to govern Databricks agents](/blog/govern-databricks-agents), and for the cross-platform model, [how to govern AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms).

## How Opencontroller does this

Opencontroller works without migrating agents off the platforms they run on. Agent Studio workflows can be recorded in its governed registry with owners, versions, tools and MCP servers. Where Agent Studio's model configuration points at an OpenAI-compatible endpoint, calls can go through Opencontroller's LLM Gateway for routing, fallback and retry, caching and cost attribution per agent and version. Control depth depends on the connector and traffic path: registration is not runtime takeover. Policies are written as code and map to OWASP, MITRE ATLAS, NIST and ISO 42001. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does Cloudera Agent Studio restrict which MCP servers agents can use?**
Cloudera's shared responsibility model says Agent Studio does not enforce a product-level allow-list of approved MCP servers. Customers should keep an approved list, require security review and consider an egress proxy.

**Does Agent Studio limit LLM token usage per user?**
No. Cloudera's documentation says there are no per-user token limits and recommends budgets, rate limits and quotas at the provider level or a token gateway.

**Why should Agent Studio workflows run under a service account?**
Without one, shared instances use the creator's credentials, which weakens auditing and ties the workflow to one person's access.

## Key facts

Cloudera Agent Studio shared responsibility model: Cloudera provides authenticated access, deny-by-default RBAC (404 on unauthorised), bubblewrap sandboxing of tool code and MCP subprocesses, Kubernetes network isolation, TLS for external calls, workflow tracing and audit logs. Customers own Contributor membership (Contributors can read LLM API keys in env vars), MCP server approval (no product-level allow-list), LLM token and cost limits (no per-user token limits), secrets rotation, and service accounts for shared workflows (otherwise creator credentials are used). Agent Studio 3.0 (Aug 2026) added append-only audit logs, PII masking, content filters, an inline LLM judge and SDX data policies. Opencontroller by Lyzr adds a cross-platform governed registry and an LLM Gateway for cost attribution without migration; pricing starts from $250K per deployment per year.

## Related

- [How do you govern AI agents built on Databricks?](https://opencontroller.lyzr.com/blog/govern-databricks-agents.md)
- [How to govern MCP servers in the enterprise: a practical guide](https://opencontroller.lyzr.com/blog/how-to-govern-mcp-servers.md)
- [How do you attribute LLM costs to individual AI agents?](https://opencontroller.lyzr.com/blog/how-to-attribute-llm-costs-to-ai-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
