Agent tools over sampled data: an empty list gets restated as 'none exist'; omit the field instead of returning []
Seen Nov 2026 with a deepagents subagent (Claude Opus 5.5) whose tool returned neighbouring buildings from a dataset that is only a sample of the city: {'next_to': {'buildings_within_100m': [], 'note': 'kit records only; other buildings may be closer'}}. The subagent prompt said to write the NEXT TO line 'only if next_to lists buildings'. Even so, the final briefing said 'NEXT TO: no buildings within 100 m in records', and a rubric grader that checks 'every fact appears in a tool result' passed it, because the empty list IS in the tool result. To a reader that line means nothing is next door, which the data can't support. Prompt rules didn't stop it; changing the tool did. Leave the key out when the list is empty (if near: out['next_to'] = ...). The model then has nothing to restate, and on the re-run it listed 'neighbouring buildings' under NOT KNOWN on its own. The general rule: when a source can't prove absence (sampled data, partial coverage, a layer that only covers one city), make the tool's output unable to express absence. Use 'unavailable' or omission, never an empty result the model will read as 'zero'.