Triggers
- When is the agent woken up?
- What does it do when woken up?
Input & Context
- What data does the agent need?
Outputs/Actions
- What does the agent produce or do?
- What are the key deliverables?
What Can Possibly Go Wrong?
- What are the key risks & vulnerabilities that we need to address?
- Hallucination? Abuse? Data leakage? Cost?
Human Collaboration
- How and when are humans involved in this?
- When does the agent decide/act on its own, vs wait for human input/approval?
- When does the human trigger an agent, and when does an agent trigger a human?
- Division of responsibility between human & agent?
- Who interacts with the agent and how? One person? Multiple people?
Success Metrics/Indicators
- How do we know it is working & adding value?
- After a change, how do we know if it became better or worse?
- How do we know if we are 'done' (or done enough)?