Salesforce Decode
Salesforcedecode
Back to questions
AgentforceAdvancedevaluationtestingquality

Implement evaluation harness with golden transcripts before agent launch

Real World Scenario

Leadership demands go-live Friday; QA has no automated way to verify 200 critical intents still work after prompt tweak.

Expected Answer

• Curate golden set: user utterance, expected topic, required citations, forbidden phrases • Automated regression scoring containment and action correctness • Block release if score drops below threshold on golden set • Include adversarial and edge cases not only happy path • Human review sample of failed regression cases • Track metrics over time not one-time pre-launch • Version golden set with agent metadata releases

Follow-Up Questions & Answers

Click to expand — each follow-up includes a direct, interview-ready answer

Direct answer: Curate golden set: user utterance, expected topic, required citations, forbidden phrases Also consider: Automated regression scoring containment and action correctness In practice: Block release if score drops below threshold on golden set Optimize for scale and operational observability.

Architect Perspective

Agents without eval harness regress silently every prompt edit — treat eval as CI gate.