AI Insights · Oct 22, 2026 · Verified
How teams test AI agents now
Teams start with 20 to 50 tasks taken from real failures and use exact checks before AI judges. When an AI does judge, it compares two answers side by side and never grades its own model.
Claim
Evals are code in the repo, not a console feature. OpenAI shut its hosted Evals product and points to Promptfoo. The recipe is small, real, graded deterministically where possible, with LLM judges used sparingly and carefully. Tracing uses OpenTelemetry GenAI conventions even though they are not stable.
Evidence
- Anthropic (2026-01-09): capability evals start at a low pass rate and graduate into regression evals near 100%; grader types code, model, human with deterministic preferred; one isolated judge per rubric dimension with an "Unknown" way out; "20 to 50 simple tasks drawn from real failures is a great start"; isolate the harness so trials do not share state; "Read the transcripts!".
- LangChain (2026-03-10): binary or low-precision scores are more reliable than numeric scales; randomise order for position bias; use a different judge model to avoid self-enhancement; turn human corrections into few-shot examples. Claude Code workflows guidance: pairwise beats absolute, fresh agent per comparison.
- OpenTelemetry GenAI semantic conventions live in a separate repo with no
tagged release and "Development" status; operation names include
invoke_agent,execute_tool,chat,embeddings. Vercel AI SDK 7 emits them; Langfuse ingests OTLP and maps GenAI, OpenInference and MLflow attributes; Claude Code exports per-user token and cost metrics via OTel. - ClickHouse acquired Langfuse (2026-01-16); it stays MIT and self-hostable.
Open questions
- Agreement targets for judges (Cohen's kappa thresholds) appear only in vendor guides. UNVERIFIED as consensus.
Sources
- https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents — blog, Thu Jan 08
- https://www.langchain.com/articles/llm-as-a-judge — blog, Mon Mar 09
- https://github.com/open-telemetry/semantic-conventions-genai — repo, Wed Oct 21
- https://langfuse.com/docs/opentelemetry/get-started — docs, Wed Oct 21
- https://clickhouse.com/blog/clickhouse-acquires-langfuse-open-source-llm-observability — blog, Thu Jan 15
AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.