AgentOps Dataset

Generate a small, realistic JSONL dataset for the agent under evaluation. Default location: .agentops/data/smoke.jsonl (referenced from agentops.yaml). These rows are repo-side release-gate inputs: keep them reviewable and deterministic, not a full replacement for Foundry dataset management.

Step 0 - Prerequisites

1. pip install "agentops-accelerator @ git+https://github.com/Azure/agentops.git@main" if agentops is missing.
2. Run agentops eval analyze first. If it reports missing dataset columns or recommends agentops-dataset, use this skill before the first eval run.
3. If agentops.yaml does not exist, run agentops init first (the init wizard will prompt for the agent reference, project endpoint, and dataset path, then create a starter .agentops/data/smoke.jsonl).

Step 1 - Pick the columns

Read agentops.yaml (and the agent code) to figure out the agent type, then choose the row schema:

input is always the user prompt. expected is the gold answer. context is the retrieved passage(s). tool_calls is a list of {name, arguments} describing the expected tool invocations.

Step 2 - Ground the rows in the codebase

• Read the README, system prompt, tool definitions, and any sample fixtures.
• Generate 5–10 rows that exercise the agent's actual capabilities.
• If the domain is unclear, generate a tiny generic draft and clearly flag it as a placeholder.

Step 3 - Write the JSONL

One JSON object per line, no trailing commas, UTF-8:

{"input": "What is the refund policy?", "expected": "Refunds within 30 days...", "context": "Refund policy: ..."}

Save to the path referenced by dataset: in agentops.yaml (default .agentops/data/smoke.jsonl).

This file is the AgentOps source of truth. In Foundry cloud evaluation, AgentOps syncs it to a stable Foundry dataset version by default and reuses the same Foundry dataset version while the JSONL content is unchanged. If the user forces dataset_sync.mode: inline, Foundry may show generated eval-data-* backing assets in the project Data/Datasets page.

Step 4 - Sanity-check

Run a quick eval and confirm rows are picked up:

agentops eval run

Open .agentops/results/latest/report.md and confirm the row count matches.

Guardrails

• Do not invent customer data, real names, or sensitive content.
• Keep rows short - datasets are meant to be quick gates, not full QA suites.
• If the user already has a domain dataset, prefer pointing agentops.yaml at that file rather than generating new rows.
• If the user asks why Foundry shows eval-data-*, explain that those are cloud-eval backing assets from inline compatibility mode; normal cloud runs should use the stable agentops-* Foundry dataset.