Argus: Evidence Assembly for Scalable Deep Research Agents

The paper's deeper contribution is a critique of how the field has been measuring scaling progress: if redundant search paths are counted as compute investment but produce correlated, overlapping evidence, then benchmark gains from parallelism are partly illusory. Argus reframes the unit of useful work as a non-redundant evidence fragment, not a completed agent trajectory.

This is largely disconnected from recent activity in our archive, as Modelwire has no prior coverage to anchor it to. It belongs to a fast-moving cluster of work on inference-time compute allocation for multi-agent systems, a space where the central tension is whether more agents doing similar things actually improves answer quality or just inflates cost. Argus takes a position on that tension by arguing that task decomposition, not headcount, is the right lever. That argument has practical stakes for anyone building or procuring research pipelines on top of frontier models, since compute costs scale with agent calls regardless of whether those calls add new information.

Watch whether Argus's evidence-assembly framing gets adopted in follow-on benchmarks like FRAMES or HELMET over the next two quarters. If independent groups replicate the redundancy-reduction gains under those evals, the architectural claim holds; if results flatten outside the paper's own test conditions, the framing may be overfitted to its benchmark setup.