The Agent Specification Builder

Write a specification for an agent.

Purpose: [ONE SENTENCE]
Users: [WHO INVOKES IT, AND HOW]
Trigger: [USER ACTION / SCHEDULE / EVENT / API CALL]
Systems it must reach: [LIST]
Constraints: [TIME, COST, COMPLIANCE, DATA]

Produce a specification covering:
1. Assignment — what it does, and a checkable definition of "done"
2. Explicit non-goals — what it must NOT do, and why each is
   excluded
3. Tool surface — the capabilities it needs, described
   functionally, not as API endpoints
4. Inputs it receives and outputs it produces
5. State: what persists within a run, across runs, across users
6. Termination conditions: success, failure, timeout, budget
   exhaustion, human intervention
7. Escalation: what conditions require a human, and how it hands
   over
8. Observability: what must be logged for someone to debug a bad
   run three weeks later
9. Blast radius: worst realistic outcome if it behaves badly, and
   what limits it
10. Open questions I need to answer before building

Do not name models, SDKs, or prices. Flag platform-dependent
decisions rather than assuming a platform.