The product is a new highest-effort mode for Exa Agent, and Exa says it is intended for deep research, list-building, and entity enrichment where completeness outweighs latency or cost.
Exa Agent Ultra frames exhaustive list construction as a distinct agent workload, rather than a variant of conversational search. The mechanism is task decomposition into subagents researching multiple domains in parallel, combined with a mix of frontier and cost-effective models chosen for the task. This matters now because Exa is attaching explicit benchmark and per-task-cost claims to that orchestration layer: it reports 81.4% soft recall on WANDR, 58.9% on WideSearch, and 93.9% on DeepSearchQA.
The immediate implication is that research-agent competition is shifting toward auditable, row-level coverage for diligence, monitoring, and market mapping, not only answer quality. The evidence boundary is that these performance and cost comparisons come from Exa’s launch post; the company says it used published competitor results where available and ran some comparisons itself, while its Find-All Company benchmark is internal. Watch whether independent evaluations reproduce the reported recall-and-cost advantages on real customer workloads with stated completeness criteria. The product’s practical test will be whether teams can obtain cited, deduplicated lists that remain cheaper to review than manually assembled research.