Testing, Evaluation and Observability for Agents

Agentforce & AITesting, Evaluation and Observability for AgentsHow to prove an agent works and keep it working: why agent testing is evaluation rather than assertion, unit-testing actions, preview traces, Testing Center suites (AiEvaluationDefinition) and the newer Agentforce Studio runner, sf agent test in CI, red-teaming and guardrails, and production observability with Session Tracing, audit data, health monitoring, cost and quality KPIs.Advanced · about 90 minutesWhat you will learn

Explain why agent testing combines deterministic assertions with LLM-graded evaluation, and design a layered test strategy

Unit-test agent actions in isolation with Apex tests and the Actions REST API

Read preview and plan traces to diagnose routing, action and grounding problems

Write test specs, create AiEvaluationDefinition suites and run them with sf agent test in a CI pipeline with quality gates

Design adversarial (red-team) tests for prompt injection, data disclosure and excessive agency

Query Session Tracing and audit data in Data 360 and define production KPIs such as resolution, escalation and deflection rates

Contents

Why agent testing is different

Layer 1: testing actions in isolation

Layer 2: preview and plan tracing

Layer 3: evaluation suites with Testing Center

Agent tests in CI

Regression strategy when instructions or models change

Red-teaming and prompt-injection testing

Production observability

Quality KPIs, cost and credits

Worked example: closing the loop for Order Concierge

Best practices

Common pitfalls

Hands-on exercises

Key takeaways