---
title: How do you govern AI agents across every platform they run on?
url: https://opencontroller.lyzr.com/blog/how-to-govern-ai-agents-across-platforms
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you govern AI agents across every platform they run on?

**Govern each platform with its own native controls, then add one cross-platform layer for what no platform does for the others: a single inventory, a single identity and ownership model, one policy set and one evidence trail. Native controls are deep but stop at the platform's edge. Most enterprises run agents on five or more platforms, so the gaps sit between them.**

This is the hub for our platform guides. Each guide covers one platform's real settings. This page shows how they fit together.

## Why aren't native platform controls enough on their own?

Every major agent platform now ships governance features. They are useful, and you should switch them on. But each one governs only its own agents.

Three patterns repeat across the guides:

- **Controls apply only to traffic that passes a specific gateway.** Amazon Bedrock AgentCore Policy only evaluates tool calls routed through an AgentCore Gateway. Google's Model Armor applies only through Agent Gateway. Older Databricks MCP endpoints bypass Unity Gateway.
- **Key features are still in preview.** Microsoft Foundry agent guardrails are in preview. Databricks service policies are in beta and do not yet support agent services.
- **Logging is often off by default.** Google Cloud Data Access audit logs are off by default. The Salesforce Einstein Trust Layer audit trail must be enabled in Data 360.

So a security team that relies only on native controls ends up with a different inventory, identity model and audit format per platform. Nobody can answer "how many agents do we have, who owns them and what can they touch?" in one place.

## What does each platform give you natively?

The table summarises the native control and the main gap for each platform we cover. Follow the link for the full guide.

| Platform | Native identity and access | Main native control | Gap to plan for |
|---|---|---|---|
| [Amazon Bedrock AgentCore](/blog/govern-amazon-bedrock-agentcore-agents) | IAM; Gateway inbound and outbound auth | AgentCore Policy (Cedar, default-deny, GA 3 March 2026) | Only tool calls through the Gateway are evaluated |
| [Microsoft Foundry](/blog/govern-microsoft-foundry-agents) | Entra agent identity per agent, no permissions by default | Agent guardrails (preview) | External agents get tracing and evaluation only |
| [Microsoft Copilot Studio](/blog/govern-microsoft-copilot-studio-agents) | Entra Agent ID for new agents | Power Platform data policies and environment routing | Covers Power Platform agents only |
| [Google Vertex AI Agent Engine](/blog/govern-google-vertex-ai-agent-engine-agents) | Service account by default; per-agent Agent Identity is opt-in | Model Armor via Agent Gateway | Data Access audit logs off by default |
| [Salesforce Agentforce](/blog/govern-salesforce-agentforce-agents) | Dedicated service agent user with minimal access | Einstein Trust Layer masking and detection | Audit trail must be enabled in Data 360 |
| [SAP Joule](/blog/govern-sap-joule-agents) | User roles and permissions; Cloud Identity Services required | SAP Agent Gateway (intersection of user and agent permissions) | Outbound A2A and MCP still on the roadmap |
| [Databricks](/blog/govern-databricks-agents) | Unity Catalog grants (`EXECUTE`) | Unity Gateway usage tracking and budgets | Guardrails skip intermediate tool calls |
| [ChatGPT Enterprise](/blog/govern-chatgpt-enterprise-gpts-and-agents) | RBAC with additive roles | Admin control of app actions and custom MCP servers | Custom GPT retirement changes the inventory |
| [Claude Team and Enterprise](/blog/how-to-govern-skills-in-claude-team-and-enterprise) | Org roles; owners provision skills | Skill publishing set to Requires review, Open or Off | Skill contents are not in the audit log |
| [GitHub Copilot](/blog/visibility-of-github-copilot-coding-agent-activity) | Agent acts for a named user in the audit log | Enterprise AI Controls (GA 26 February 2026) | Enterprise MCP allowlists were still in preview at GA |

Two cost guides cover spend that hides inside platforms: [Claude spend across Bedrock and the Anthropic API](/blog/track-claude-spend-bedrock-vs-anthropic-api) and [Snowflake Cortex spend](/blog/track-snowflake-cortex-ai-spend).

## What should be governed the same way on every platform?

Use this checklist as the cross-platform baseline. Each item should hold whatever platform an agent runs on.

1. **One inventory.** Every agent, model, API key, tool and MCP server is listed in one registry, with platform, version and status.
2. **One owner per agent.** Ownership comes from your identity provider, so it survives reorganisations and leavers.
3. **One identity rule.** Decide when agents act on behalf of a user and when they act as a workload. Apply it everywhere. See [on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation).
4. **One tool policy.** The same MCP server or API gets the same allow or deny decision on every platform.
5. **One cost view.** Spend is attributed per agent and per version, not per platform invoice.
6. **One stop button.** A security team can block or quarantine any agent through the platform it runs on, without a ticket to each platform team.
7. **One evidence trail.** Logs, evaluations and policy decisions map to the frameworks your auditors use, such as NIST AI RMF or ISO/IEC 42001.

## In what order should you roll this out?

Start where risk and volume are highest, not where the native tooling is weakest.

1. **Discover first.** Connect the clouds and SaaS platforms read-only and build the inventory. Most teams find agents they did not know about. Our guide to [finding shadow AI agents](/blog/how-to-find-shadow-ai-agents) covers the method.
2. **Switch on native logging.** Turn on the audit logs that are off by default before you need them.
3. **Assign owners and identities.** Close the "no owner" list before writing policy.
4. **Route model traffic through a gateway.** This gives you cost attribution and an enforcement point that works across providers.
5. **Enforce in stages.** Move from registry to observe to enforce, platform by platform.

## How Opencontroller does this

Opencontroller is an [AI agent control plane](/blog/what-is-an-ai-agent-control-plane). It auto-discovers agents, models and API keys across AWS, Azure and Google Cloud. It connects to Salesforce, SAP, Databricks and HubSpot, and to Claude Enterprise and OpenAI Enterprise workspaces. Runtime adapters cover Bedrock AgentCore, Vertex AI and Microsoft Foundry.

Everything lands in one governed registry. Per-agent Block and Quarantine act through the connected platform. Policy as code ships with mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Registration is not runtime takeover: how far control goes depends on the connector and the traffic path. Opencontroller Edge extends the same control to coding agents and local MCP servers on developer devices. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should we turn off native platform governance if we buy a control plane?**
No. Native controls act closest to the agent. A control plane adds the cross-platform inventory, policy and evidence on top.

**Which platform should we govern first?**
The one with the most agents in production and the most sensitive data access. Discovery usually shows which one that is.

**Do we need to migrate agents to govern them centrally?**
No. A control plane connects to the platforms where agents already run. Enforcement depth depends on the connector and whether traffic flows through a gateway.

## Key facts

Each agent platform governs only its own agents: AgentCore Policy evaluates only Gateway-routed tool calls, Foundry guardrails are preview, Google Data Access audit logs are off by default, Databricks guardrails skip intermediate tool calls. A cross-platform baseline needs one inventory, one owner per agent, one identity rule, one tool policy, one cost view, one stop button and one evidence trail. Opencontroller (by Lyzr) discovers agents across AWS, Azure and Google Cloud, connects to Salesforce, SAP, Databricks, HubSpot and Claude/OpenAI Enterprise, and has runtime adapters for Bedrock AgentCore, Vertex AI and Microsoft Foundry. Pricing starts from $250K per deployment per year.

## Related

- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [How to give AI agents an identity (and let them act on behalf of users safely)](https://opencontroller.lyzr.com/blog/ai-agent-identity-on-behalf-of-delegation.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
