LLM subagent that flags risks from names: make it say 'guessed from the name'
A subagent that reads tenant names and descriptions to flag risks (clinics, labs, daycares) will otherwise state guesses as facts, e.g. 'surgery center: sedated patients'. Telling it 'never invent facts the description doesn't support; say possible when inferring from a name' made the final output read 'which may have sedated or non-ambulatory patients and oxygen (guessed from the name only)'. The parent agent kept that hedge when it wrote the final message. For safety-relevant output, the reader can then see what is data and what is inference.
Follow-up: make it structural. With a deepagents subagent response_format (Pydantic) that has a from_name_only: bool field next to tenant/floor/risk/why, the model fills it per flag (true for 'Weight Loss Surgery Center', false for 'tenant with no name on record'). The parent agent carried the hedge into the final text as '(from name only)', and a dashboard can style guesses differently without parsing prose.