DOMAIN-SPECIFIC HARNESSES

CHOOSING THE RIGHT SETTINGS FOR A DOMAIN When you design a harness for a new domain, ask the same set of questions. What could go wrong, and how severe would the harm be? Which tools does the agent truly need, and which should it never have? Which actions are irreversible and must be approved by a human? What rules or laws apply? What does correct output look like, and how will we measure it? Where should the agent run, and what happens when it fails? The answers decide the permission levels and verification steps, which is the heart of domain-specific harness design.