When search results arrive in a different order, agents seek different evidence

When an agent gets the same evidence in a different order, does it search for different documents? That question matters because an answer can look right even when its supporting evidence depends on which results arrived in time to guide an early search decision.

Confluent Evidence Acquisition coordinates subquestion reasoning and shared searches so fixed-plan agents can make decisions consistently across delivery orders.

If an agent receives the same starting evidence in a different order, can that change what it goes on to search for? The final answer alone does not show whether the agent had a reproducible path to its supporting evidence.

The researchers replayed fixed packets in different orders, then introduced CEA to coordinate reasoning around each subquestion’s assigned evidence and allocate shared searches independently of completion order.

With question-only plans, acquired evidence stayed near same-order repeat variation; accuracy gains over early commitment appeared on two tasks but not the third.

The proof assumes a fixed plan, deterministic retrieval, and response behavior that depends on the request rather than earlier execution history.

The experiments cover specific models, question sets, plans, and retrieval conditions, so they do not establish invariance for arbitrary agents or settings.