---
title: How do you govern Salesforce Agentforce agents?
url: https://opencontroller.lyzr.com/blog/govern-salesforce-agentforce-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you govern Salesforce Agentforce agents?

**Govern Agentforce agents mainly through the permissions of the user they run as. Service agents run under a dedicated agent user that starts with minimal access, so expand it only by least privilege. Add the Einstein Trust Layer's masking and audit trail, test every change in a sandbox with Testing Center, and monitor sessions with Agentforce observability. Agents outside Salesforce need their own inventory.**

Agentforce agents act inside your CRM, so the usual Salesforce security model is the core of their governance. Profiles, permission sets, field-level security and sharing rules decide what an agent can read and change. Salesforce adds AI-specific controls on top.

## Who does an Agentforce service agent run as?

A service agent runs as a dedicated user, the "EinsteinServiceAgent User", with the Einstein Agent User profile and licence. Salesforce assigns permission sets such as "Agentforce Service Agent Secure Base" and an agent-specific permissions set. The [agent user docs](https://help.salesforce.com/s/articleView?id=ai.agent_user.htm&language=en_US&type=5) say the user starts with "minimal access so that your agent is secure by default" and tell admins to expand access "based on the principle of least privilege".

In practice, the risk is permission creep. A team hits a failing action, adds a broad permission set to fix it, and never removes it. Review the agent user's effective access as you would any integration user. Our guide to [on-behalf-of delegation](/blog/ai-agent-identity-on-behalf-of-delegation) covers when an agent should act as itself and when as the person it serves.

## What does the Einstein Trust Layer do?

| Trust Layer feature | What it does | Governance use |
|---|---|---|
| Zero data retention | Third-party LLMs do not retain your data or train on it | Vendor risk evidence |
| Data masking | Masks sensitive data before the prompt reaches the LLM | Privacy control |
| Toxicity and prompt-injection detection | Scores outputs and defends prompts; some detectors are listed as beta | Runtime safety |
| Secure data retrieval | Keeps Salesforce role-based access in place when grounding | Least privilege |
| Audit trail | Prompts, responses and trust signals stored in Data 360 (formerly Data Cloud) | Evidence and investigations |

The audit trail is not automatic. You have to turn on data collection, and it needs Enterprise, Performance or Unlimited edition plus an Einstein add-on. Check that it is on before you rely on it in an [evidence pack](/blog/regulator-inspection-evidence-pack-for-ai-agents).

## How do you test changes safely?

Agent behaviour is shaped by subagents (called topics until April 2026), their instructions and their actions. Any change can alter what the agent does. Testing Center runs batch tests, can generate test cases, and evaluates subagent recognition, action execution and knowledge retrieval. Salesforce warns that tests can modify CRM data, so run them in a sandbox only.

## How do you monitor agents in production?

Agentforce observability (formerly Command Center) gives session-level tracing, health monitoring, near-real-time alerts and quality scores. The session trace data model records inputs, responses, reasoning steps, LLM calls and guardrail checks in Data 360. A beta OpenTelemetry export API can send traces to tools such as Splunk, Datadog or New Relic, with limits of one session per query and the last 72 hours.

## What should an Agentforce governance checklist include?

1. Each agent has a named business owner and a technical owner.
2. The agent user's permission sets are reviewed and justified line by line.
3. Field-level security hides fields the agent never needs.
4. Trust Layer data masking is configured for your sensitive fields.
5. Audit trail data collection is turned on and retained per policy.
6. Every subagent, instruction or action change passes Testing Center in a sandbox.
7. Observability alerts route to a team that can act on them.
8. A tested way exists to deactivate the agent quickly.

## What does Agentforce not cover?

The Trust Layer, Testing Center and observability govern agents that run on Agentforce. Your estate likely also includes agents on AWS, Azure, Google Cloud, SAP, Databricks and developer machines. Salesforce sells MuleSoft Agent Fabric separately for agents built elsewhere. Whatever you choose, aim for one inventory and one policy set across platforms. That is the role of an [AI agent control plane](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller connects to Salesforce alongside AWS, Azure, Google Cloud, SAP, Databricks, HubSpot and Claude and OpenAI Enterprise workspaces. Agentforce agents appear in the same governed registry as the rest of your estate. Posture flags risky permissions and configuration. Per-agent Block and Quarantine controls act through the connected platform, and rogue-agent handling can alert, quarantine or kill on behavioural deviation. Enforcement depth depends on the connector and traffic path. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What permissions does an Agentforce service agent have by default?**
It runs as a dedicated agent user that starts with minimal access. Admins add permission sets, and Salesforce advises doing so by least privilege.

**Where are Agentforce prompts and responses logged?**
The Einstein Trust Layer audit trail stores prompts, responses and trust signals in Data 360, once data collection is turned on.

**Can Testing Center change real data?**
Yes. Salesforce warns that tests can modify CRM data, so run Testing Center in a sandbox.

## Key facts

Agentforce service agents run as a dedicated EinsteinServiceAgent user that starts with minimal access; expand by least privilege. The Einstein Trust Layer provides zero data retention with third-party LLMs, data masking, toxicity/prompt-injection detection and an audit trail stored in Data 360 (must be enabled). Topics were renamed subagents in April 2026. Testing Center can modify CRM data, so use a sandbox. Opencontroller (by Lyzr) connects to Salesforce and other agent platforms. 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 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)
- [What evidence will a regulator ask for about your AI agents?](https://opencontroller.lyzr.com/blog/regulator-inspection-evidence-pack-for-ai-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
