Exa has introduced a new high-effort mode for its Exa Agent API called Agent Ultra. This tool is designed specifically for deep research tasks, primarily focusing on exhaustive list building and entity enrichment. By deploying a swarm of subagents, the API navigates thousands of information sources to compile data. MarkTechPost reports that this release targets professional-grade data gathering where standard search queries might fail to capture the full breadth of a specific domain.
Performance and Benchmark Results
The release of Agent Ultra positions Exa as a direct competitor to high-end research agents from other major AI developers. According to the reported data, Agent Ultra outperformed several prominent models on four distinct benchmarks. These competitors include Opus 5.5, GPT-6 Astra, and the Perplexity Agent. A notable performance metric cited is an 81.4% soft recall on the WANDR benchmark. This suggests that the subagent swarm architecture is optimized for retrieval accuracy and completeness, particularly in the context of list building—a task that requires identifying every relevant entity within a defined set.
Operational Unknowns
While the benchmarks indicate strong performance, several operational details remain undisclosed. The announcement did not specify the exact pricing structure for this mode or whether there are rate limits associated with the subagent swarm. It is also unclear what specific compute resources are required to power these extensive searches or the average latency for a single deep research request. Furthermore, the report did not provide a full list of all four benchmarks used, nor did it detail the specific entity enrichment capabilities beyond basic data retrieval. The technical requirements for integrating the API into existing workflows also remain unstated.
As AI agents move toward more autonomous research workflows, Agent Ultra represents an attempt to automate the most labor-intensive aspects of data collection. Its long-term utility will likely depend on how effectively it handles niche data sources in real-world applications.
Lead image: Brett Sayles / Pexels
