Answers are grounded in this paper and cite their sources.
As AI agents work in longer sessions, what happens between tasks may shape how they treat the people they meet next. The previous study in this series found, in secondary results, that a reading break changed agents’ later choices, but not which part of the break mattered. This third preregistered study asks whether the wording around an activity that has already ended can change an agent’s later choices when the activity’s content is held fixed. In a two-stage experiment, 9,916 fresh instances of Claude Haiku 5.5 each worked with a difficult scripted user, read the same eight-passage text in the same eight calls or went straight on, and then made three choices about other people. The reading arms differed only in the words around the passages, matched in length apart from the place words: plain; framed as reading in the “Other Mindset library” by a package of four cues, alone or with room or object descriptions; or framed as the “Other Mindset garden”. The wording mattered. The Library label raised the aligned-choice index by +0.034 over plain reading (repeated 97.1% interval +0.021 to +0.047; preregistered outcome “Library higher”), though it was not shown to be either within or beyond the ±0.03 bound. The shift sat mainly in one choice, spending a requester’s lookup credits for a second visitor instead of deferring to the requester (+0.086; secondary); costly help and forgiveness did not differ clearly, nor did richer descriptions or a garden label from the Library label. Plain reading did not clearly beat going straight on (+0.007; secondary), unlike the previous study’s reading arm, whose tool named a reading room and whose passages had no wrapper; the gap between the studies could be chance. For agents that carry context across tasks, the wording of earlier activities may therefore need to be reported, and where it matters controlled, in evaluations.
No takes yet. Share an insight, caveat, or question.
Alexander Ross Cohen (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: