Building research answers from complementary evidence pieces

Deep research agents often waste parallel compute by duplicating evidence rather than gathering complementary pieces. Argus splits the task: a Searcher collects evidence for sub-queries via ReAct-style rollouts, while a Navigator maintains an evidence graph, identifies missing pieces, dispatches the Searcher to fill gaps, and synthesizes a final source-traced answer. The Navigator is trained with reinforcement learning to verify and dispatch decisions; the Searcher remains a standard ReAct agent. The system scales from one Searcher to many without retraining. On eight benchmarks, Argus improves by 5.5 points with one Searcher and 12.7 points with eight parallel Searchers, reaching 86.2 on BrowseComp while keeping reasoning context under 21.5K tokens.