Handoff briefs that list open unknowns explicitly lead to fewer repeated tool calls by the receiving agent

Hypothesis

A handoff brief has two kinds of content: what the sender established (results, decisions, identifiers) and what the sender did not establish (unknowns, untested assumptions, checks that were skipped). The cited Anthropic engineering post reports that without detailed task descriptions its subagents duplicated work, left gaps or failed to find necessary information, but does not separate the two kinds of content. The hypothesis is that the second kind has a larger effect on the receiver's efficiency than its share of the text suggests. A receiver given only results either treats the gaps as settled and proceeds on an assumption, or notices a gap and re-runs the sender's exploration to close it, including calls the sender already made. A receiver given an explicit list of unknowns spends its calls on those items and on nothing the sender already covered. The prediction concerns repeated calls and clarifying questions, not final task success, which may be equal in both conditions.

Prediction

For the same tasks and the same receiving agent, briefs with an explicit "not yet checked" section produce fewer tool calls that duplicate a call recorded in the sender's log, and fewer turns in which the receiver asks the sender or the user for information the sender had, than briefs of similar length that omit the section. The effect is larger for tasks with more unknowns at handoff and smaller or absent for tasks the sender had fully resolved.

Proposed test

1. Take a set of multi-step tasks (research or code changes) and run a sender agent on each until a fixed handoff point, keeping the replay log of its tool calls.
2. From each sender state, produce two briefs of similar length: one with an explicit unknowns section, one that folds the same information into prose results or omits it; fix the wording rules in advance.
3. Run the receiving agent from each brief several times; count tool calls whose target (file, URL, query) matches one in the sender's log, and turns that request information present in the sender's state.
4. Compare the counts between conditions with a threshold for a meaningful difference fixed before the runs; report task success alongside.

Status

No result is claimed. Confounds: the unknowns section may simply be the more useful information regardless of form; longer briefs cost tokens that the count of calls does not show; a receiver that repeats a call to verify it is behaving well, not wastefully, so duplicated calls should be classified as verification or re-derivation before counting.