Agent definition card

Complete this card before creating an agent. If a field cannot be answered clearly, its boundary has usually not been fully designed.

Field	What to enter
Agent name	Name it after its responsibility domain; avoid using only a persona name
Business objective	Whose problem it solves and what that problem is
Success criteria	How to determine that the work is complete and correct
Allowed inputs	Which tasks, context, and identities it accepts
Allowed outputs	What structure and evidence it must return
Allowed tools	Which tools it may invoke and whether they have read or write access
Data boundary	Which data it may access
Permission boundary	Which actions are forbidden or require approval
State	Which task state must be persisted
Budget	Limits on time, tokens, cost, and retries
Termination conditions	Conditions for success, failure, handoff to a human, and refusal
Evaluation	Which datasets and metrics are used for validation
Owner	Who maintains and approves it, and responds to incidents

Minimal example

name: refund-analysis-agent
goal: Identify the main drivers of refund-rate changes within the specified scope
success:
  - Every conclusion is supported by query evidence
  - The denominator, time range, and filters are explicit
allowed_tools:
  - metadata_search
  - read_only_sql
forbidden:
  - Write to production databases
  - Bypass row- and column-level permissions
stop:
  - Evidence meets the release threshold
  - Budget is exhausted
  - Permissions are insufficient
  - Human clarification is required