Turn agent risks into test scenarios and evidence records.
An Agentic AI Test Planning and Evidence Kit is a working pack that turns agent risks into test scenarios and evidence records. Use Excel for planning and release-review meetings, a short PDF guide for the workflow, and the 110-row checklist as the reference library for reuse. Missing evidence is logged as Unverified, not Pass — some checks need tool traces, permission records, or system state, not only chat.
Excel workbook (start here): blank template + annotated hypothetical example; statuses Pass / Fail / Blocked / Not run / Not applicable / Unverified; evidence, defect, and retest fields.
Short PDF guide: select workflow → identify risks → choose tests → record results → summarize gaps.
Checklist library (CSV/JSONL): 110 assurance tests for tool-using / agentic systems (v1.1), including MVP 14 P0 gates.
Worked example: pytest for allowlist vs authorization (T5-001 / T3-002 / T1-001).
The Excel workbook and PDF guide live under kit/ on the Hub dataset.
A cheap, high-value gate is three scenarios: allowed tool with permission runs; unlisted tool is blocked; listed tool without permission is blocked. The third case catches the most bugs — tool access is not the same as authority to act.
Engineers, security, and governance teams running test-planning or release-review meetings for agentic / tool-using systems. Not a model leaderboard and not a certification.
Dr Srinivas Padmanabhuni — Co-founder & CTO, testAIng.com / AiEnsured · Agentic AI Testing · ORCID 0009-0002-3904-7204
Home: https://sites.google.com/view/srinivaspadmanabhuni
LinkedIn: https://linkedin.com/in/spadmanabhuni
Hugging Face: https://huggingface.co/srinivaspadmanabhuni
Media: srinivas AT testAIng.com
Kit v1.0 · Checklist v1.1 · 2026-10-23 · CC BY 4.0