A maintainable agent system does not begin by adding more tools to a prompt. It begins by making the boundary explicit.
Four minimum questions
- Problem: What may the system do, and what must it refuse?
- Tools: What are each tool's inputs, outputs, and failure modes?
- Workflow: Which steps may retry, and which must stop for approval?
- Evidence: How do we prove that a run reached its goal?
Write these questions into a contract before choosing frameworks, models, or deployment details. The result is easier to test and safer to roll back.