Index, Formulate, or Resolve? How External Support Changes What Agents Do

When agents make decisions from historical records, retrieving evidence is only the start. They must locate relevant records, formulate and resolve the relation questions needed to use them, then calculate, construct, and submit a usable plan. Across three progressively finer-grained studies, we compare ways of dividing this work by organizing or completing different parts before execution. Study 1 asks whether explicitly identifying which historical records matter and which instructions still apply improves the final artifact, and whether any benefit grows when less interpretation is supplied; the expected gain is not established. Study 2 separates indexing, organizing unresolved questions, and supplying intermediate results. Study 3 further separates reference recovery from judgments about current rules and status. Across all three studies, source evidence remains complete, and agents still have to calculate, construct, and submit their own artifacts. In Study 3, adding correct rule judgments after reference recovery lowers mean provider-reported reasoning in seven of eight configurations; four history-resampling intervals lie entirely below zero. DeepSeek Pro is the only exception: one retained budget-exhausted run flips its overall JUDGE–GROUND mean positive. For Qwen Flash, adding reference results to the unresolved worksheet raises full delivery from 36/60 to 46/60. Further judgments yield its lowest mean reasoning and actor cost, but no additional delivery gain. Relative to sources alone, complete support lowers both reported reasoning and actor cost for Flash, whereas both OpenAI configurations also use less reported reasoning but cost more as cumulative input grows. Within each paired model family, the delivery gap narrows under complete support but remains. External support should therefore be evaluated by what work it completes, which model receives it, and which outcome matters. These are actor-side results under oracle-correct support; support production is not measured.