The Atlas Research coordinator writes its complete subtask list for a research question before any subagent runs, then dispatches that list stage by stage without revising it. On questions whose shape is known in advance the reports score well. On 45 open questions where the first web search turns up an unexpected regulatory ruling, the finished report covers the originally planned lines of enquiry thoroughly and investigates the ruling nowhere, even though every subagent completed successfully and the coordinator's running summary mentions the ruling in passing. What best explains this?
- AThe subtask list is dispatched stage by stage, so the ruling arrived after its own stage had already been consumed, and dispatching all remaining subtasks in a single batch instead would let the later stages see it.
- BA decomposition fixed before any evidence exists can only encode the lines of enquiry that were knowable at planning time, so the coordinator has to re-plan against each stage's findings and mint subtasks the original plan could not have contained. Correct
- CEach subagent returns its own findings without a judgement about their significance, so requiring every subagent to rate the importance of what it found will cause the ruling to be picked up and pursued.
- DThe planning step ran without plan mode, so the coordinator produced an execution list rather than a reviewed plan, and enabling plan mode for the planning call will make the resulting decomposition cover unanticipated findings.
Why A is wrong: It correctly notices that timing matters, but batching the remaining subtasks commits to the plan sooner rather than later, which makes the same problem worse rather than better.
Why B is correct: Correct. Adaptive decomposition treats the plan as revisable state updated by results, which is the only way a branch nobody anticipated can enter the workflow.
Why C is wrong: Self-reported significance is superficially attractive as a routing signal, but a rating changes nothing when no mechanism exists to create a new subtask in response, and self-assessed importance is a weak proxy for research value.
Why D is wrong: Plan mode is a real feature and naming it lends the option credibility, but it constrains an agent from acting while it plans and cannot give a planner evidence that has not been gathered yet.