Agentforce testing and governance
Test scenarios, guardrails, human handoff, monitoring and change control for Agentforce agents, using the tools and statuses Salesforce publishes.
An agent that passed a demo can still fail in production. It might answer outside its scope, miss an escalation or drift after a prompt change. Our Agentforce testing and governance service gives you repeatable tests, clear guardrails, a safe handoff to people and a routine for monitoring and changing agents after launch.
For the testing techniques themselves, read Agentforce agent testing and guardrails. To build agents, see Agentforce and AI automation.
What's included
| Area | What we set up | Salesforce tooling and status |
|---|---|---|
| Test design | Scenario sets for normal requests, edge cases, refusals and escalations | Agentforce Builder, where Salesforce says agents are built, tested and monitored |
| Scoring | Custom measures of answer quality | Custom Scorers: API GA, UI Beta |
| Prompt testing | Batch tests of prompt templates | Prompt Builder batch prompt testing: Beta |
| Voice | Tests for voice conversations | Voice and Conversation Testing: Beta; Voice Observability: GA |
| Experiments | A/B tests between agent versions | A/B Experimentation for agents: Beta |
| Guardrails | Scope, data masking, toxicity checks and permissions | Einstein Trust Layer |
| Monitoring | Health dashboards and conversation review | Agent Health Monitoring: GA; Agentforce Observability |
| Optimisation | Regular review of failing conversations | Agent Optimizer: GA announced for October 2026 |
How we work
- 01
Discovery
Review each agent's purpose, topics, actions, data access and current test coverage.
- 02
Design
Write test scenarios and pass criteria, define guardrails and escalation rules, and agree owners.
- 03
Build
Set up test sets, scorers, monitoring and the review routine, and document the release process.
- 04
Test
Run the full test set, fix gaps, and re-run before any change goes live.
- 05
Launch
Release in stages, with monitoring and a clear rollback path.
- 06
Support
Review conversations and scores regularly, update tests as topics change, and report to owners.
Human handoff
Agent Script, used by the new Agentforce Builder, includes an escalate instruction, so handoff can be designed into the agent's logic rather than added later. We define when an agent must hand over, what context passes to the person, and how that is tested. For a worked service example, see the service agent with human handoff.
Governance
Good governance is mostly ownership and routine. Each agent needs a named owner, an approval step for changes, a test run before every release and a regular review of conversations. Lifecycle Center (formerly DevOps Center) adds governance for agent-generated code, which Salesforce describes as GA. Observability Center in Salesforce Guardian is in Beta. Our responsible AI page explains how we approach this work.
Prerequisites and licences
You need Agentforce enabled in your org and users with the Manage AI Agents permission to work in Agentforce Studio. Some monitoring and security tools are separate add-ons, so we confirm what your licences include.
Common mistakes
- Testing only the happy path.
- Changing prompts or topics in production without re-running tests.
- No owner after launch, so nobody reads the conversations.
Related services
Salesforce AI readiness assessment
Check data, process and licence readiness first
→Salesforce testing and QA
Testing across the wider org
→Salesforce security and governance
Permissions, access and platform controls
→Questions we hear
Can we test agents before customers see them?
How many test scenarios do we need?
Should we rely on Beta testing features?
Who should own an agent after launch?
Sources
Make your agents testable
Tell us which agents you run or plan to launch. We'll suggest a test and governance setup that fits.
