Positions

Position papers from the ADPS community on contested questions in agent engineering. Each takes a clear, evidence-led stance — what we hold, why, and where the boundary is.

• When to Use Multi-Agent — and When Not To
  Multi-agent is not the next frontier — it is one bounded design choice in the matrix; single-agent plus tools plus a Harness covers 70%+ of production.
• Where Prompt Engineering Fits
  Prompt is one input to the Harness, not architecture — it is 1/10 of agent engineering, not the center.
• Governance and Safety
  Governance is not a bolt-on; it is part of design — OWASP Agentic Top 10 mapped to the eight principles.
• Evaluation as Engineering
  Evaluation is design itself — pre-design, pre-prod, production; miss one stage and the system drifts.