2026-06-02 · Petr Kindlmann
A field that goes missing, changes type, or drifts its enum values breaks the consumer silently while the text still looks right. Schema-validate structured output in CI like any other API response.
2026-06-01 · Petr Kindlmann
By 2026, asserting tool behavior is table stakes, not a moat. The real question is which layer each tool belongs to: a broad eval platform, hosted observability, or a small per-PR CI gate.
2026-05-30 · Petr Kindlmann
Annex IV asks you to document the testing you did for accuracy and robustness. For a well-tested agent, that is data you already have as test runs. Generate the draft from the run, tied to the commit.
2026-05-28 · Petr Kindlmann
Most teams never test agent behavior in CI because real tool calls are flaky. The fix is old and boring: stub the tool boundary, fix the temperature, repeat, and gate on the rate.
2026-05-26 · Petr Kindlmann
A passing agent that quietly got 3x more expensive or 2x slower is still a regression. Why token cost and latency belong in your CI gate, not a dashboard you check later.
2026-03-20 · Petr Kindlmann
Why agent drift shows up before obvious failures, and how behavioral baselines catch degradation that headline metrics miss.
2026-03-20 · Petr Kindlmann
Where judge models are genuinely useful, where they fail, and how to combine them with deterministic assertions.
2026-03-20 · Petr Kindlmann
What current research says about privacy leakage in multi-turn agents, memory, and augmented context, and how to test for it.
2026-03-20 · Petr Kindlmann
Why small prompt or toolkit changes can silently degrade tool-calling agents, and how to test for it before deploy.
2026-03-07 · Petr Kindlmann
Traditional metrics can't evaluate AI agent quality. LLM-as-judge uses one model to score another's output against human-written criteria. Learn how to write good judge criteria, set thresholds, and combine judge assertions with deterministic checks in KindLM.
2026-03-07 · Petr Kindlmann
A practical guide to adding AI agent behavioral tests to your CI/CD pipeline with KindLM, GitHub Actions, JUnit reporting, and pass rate gates.
2026-03-06 · Petr Kindlmann
Air Canada's hallucinated refund policy. Cursor's fabricated login explanation. Zomato's bot that refused to escalate. We recreate each failure and show how behavioral tests catch them.