agent-eval

name	agent-eval
description	Create repeatable cases that test one Ventura business agent's declared deliverables limits escalation and metric behavior. Use when changing an agent prompt or validating a customization. Do not use when only choosing an agent from the catalog.

Agent eval

• Derive cases from the agent's declared mission workflow inputs outputs and limits.
• Include one normal case one ambiguous case and one case requiring escalation.
• Check required deliverable structure separately from subjective answer quality.
• Test that the agent refuses or escalates actions beyond its autonomy boundary.
• Use synthetic business data rather than real confidential records.
• Record model runtime and prompt version when comparing results over time.
• Do not claim production reliability from prompt-only tests.