Positions

ADPS position papers address contested questions in agent engineering. Each states a position, supporting evidence, and an engineering boundary.

• When to Use Multi-Agent — and When Not To
  Multi-agent is a bounded design choice in the matrix. Establish a comparable production baseline with a single agent, tools, and a Harness first.
• Where Prompt Engineering Fits
  Prompts control the language layer of a model call; Harness state, tools, permissions, and validation define the system boundary.
• Governance and Safety
  Governance belongs in runtime design: permissions, auditability, blast radius, and autonomy changes require explicit controls.
• Evaluation as Engineering
  Evaluation spans design, pre-production validation, and production operation, providing evidence for architecture, release, and rollback decisions.