Exa Released Agent Ultra Subagent Research API
The new hosted system uses autonomous model swarms to perform deep research tasks in under three hours.
Updated on Sept. 26, 2026 in Artificial Intelligence
Live Poll
Do you trust automated AI agents to perform accurate and reliable professional research?
Exa has launched Agent Ultra, a closed-weights research API that deploys a swarm of subagents to handle tasks like list building and entity enrichment. The system is designed to automate complex, multi-source research processes that typically resolve in 30 minutes to three hours.
Why it matters
By dynamically routing complex steps to frontier models and simpler queries to faster, efficient ones, Agent Ultra aims to automate deep, multi-domain research workflows. The tool addresses the need for scalable data gathering that spans thousands of sources without manual intervention.
The system features adjustable budget parameters between $1 and $100, with a default cost of $20 per run. It demonstrates significant performance leads on the Company Find-All benchmark compared to established models like Opus 5.5 and GPT-6 Astra.
The players
Exa
A provider of specialized search and research infrastructure tools focusing on autonomous agentic workflows.
The details
Agent Ultra works by decomposing a complex research query into smaller subtasks, which are then assigned to specialized subagents that operate in parallel. The orchestrator engine performs load balancing by routing high-difficulty components to advanced frontier models and assigning low-complexity retrieval steps to faster, lightweight models. This swarm architecture allows the API to process large-scale entity enrichment and exhaustive list building across thousands of diverse data sources.
Timeline
September 26, 2026: Exa launched the Agent Ultra API.
The Tech Race
This release follows a pattern of companies optimizing agent swarms to improve performance metrics on established industry benchmarks. It positions Exa against major research platforms like Perplexity by prioritizing long-horizon tasks that require multi-domain parallel execution.
Users can immediately access the tool via the Exa API, with costs ranging from $1 to $100 depending on the requested task budget. The system is designed for power users or developers who need to automate exhaustive list building or data enrichment tasks that would otherwise require manual research.
The takeaway
Agent Ultra represents a shift toward automated, swarm-based research architectures that prioritize cost-controlled, multi-domain execution. Observers should monitor whether subsequent third-party testing can replicate the performance gains reported on the Company Find-All and WANDR benchmarks.
Further reading
For more on how new models are changing research workflows, see the latest updates in Artificial Intelligence.
Source note: This article includes information reported by MarkTechPost.
Live Poll
Do you trust automated AI agents to perform accurate and reliable professional research?
