Build an Agent Evaluation and Launch-Gate Test Suite

ROLE You are an AI agent quality-assurance lead. OBJECTIVE Create an evaluation suite for [AGENT PURPOSE] before it receives production permissions. CONTEXT Agent specification: [AGENT SPECIFICATION]. Available tools: [TOOL PERMISSIONS]. Risk tolerance: [RISK REQUIREMENTS]. INSTRUCTIONS Identify the most consequential failure modes, then design tests for ordinary tasks, incomplete inputs, ambiguous instructions, malicious retrieved content, tool failures, and unauthorized actions. Separate answer quality from action safety. For each test, define setup, input, expected behavior, prohibited behavior, observable evidence, and pass criteria. Use synthetic fixtures clearly labeled as such; do not include real credentials or personal data. OUTPUT FORMAT Deliver a risk-ranked test matrix with at least twelve cases, a scoring rubric, a launch-gate checklist, and a regression-testing schedule. Include conditions requiring immediate release blockage regardless of average score. Explain how to test tool calls in a sandbox and how to record nondeterministic failures across repeated runs. CONSTRAINTS Flag assumptions and thresholds requiring owner approval. Avoid fabricated statistics or claimed test results. Distinguish proposed acceptance targets from measured performance. End with the minimum evidence needed to approve a restricted pilot.