Writing Tasks

Task definitions live in tasks.json.

The starter fixture is examples/calendar-email/tasks.json. Each task includes:

• id
• prompt
• expectedTool
• optional successCriteria

Use expectedTool to name the tool the agent should choose. Use none when no tool should be selected.

Good eval sets include:

• clear tool-use tasks
• no-tool tasks
• ambiguous tasks that expose overlap
• unclear tasks that show whether the agent should ask for clarification
• intentionally tricky tasks that protect against regressions

Tags are not a first-class task field yet. Use clear id values and successCriteria until a future schema adds tagging.

Example:

{
  "id": "email-status-update",
  "prompt": "Email Jordan a short status update about the release.",
  "expectedTool": "send_email"
}