NVIDIA releases AI-Q for deep research in agent frameworks

NVIDIA has introduced AI-Q, a system that adds specialized deep-research skills to agent frameworks. Instead of building research capabilities into general-purpose agents, AI-Q moves that function into a separate backend, letting orchestrators such as Claude Code focus on managing workflows.

Why ordinary agents fall short

Frameworks such as Claude Code, Codex and LangChain Deep Agents work well as orchestrators: they manage sessions and tool chains, execute code and understand the developer's intent. But when it comes to serious research, such as synthesizing many documents, producing well-founded reviews and analysis with cited sources, complexity shoots up. Building that logic into an agent is inefficient and awkward.

How AI-Q works

AI-Q works as a standalone skill, an extra capability the agent can call. The system includes a full research pipeline: intent classification, a dialogue with the user for clarification, shallow search for quick answers and deep analysis for multi-point synthesis.

• Query classification: decide whether a quick lookup or full research is needed
• Clarification with a human: if the question is unclear, the system asks follow-up questions
• Shallow search: quickly find an answer to simple queries
• Deep analysis: synthesize many sources and provide links

Quality is assessed against standards such as FreshQA and Deep Research Bench. NVIDIA uses real benchmarks to make sure the system delivers reliable results.

Enterprise data is built in

A notable point: AI-Q supports authenticated MCP (Model Context Protocol) servers as data sources. This means the agent can research a company's internal documents without leaking them outside. NVIDIA provides three authentication patterns: open servers, service accounts for shared enterprise data, and bearer tokens to preserve the user's identity.

"The full research pipeline, including classification, clarification,

shallow and deep analysis, is exposed as a high-level capability"

Flexible deployment

Organizations can deploy AI-Q on their own hardware: Docker Compose on a developer's laptop or Helm in a regulated data center. Sensitive source data stays inside; only cited results go out. This matters for companies that do not want to send internal documents to the cloud.

What this means

AI-Q reflects a 2026 trend: specialized components instead of universality. Agents get better when they can delegate complex tasks to narrow tools. For developers, this means deep research is now built not into the LLM itself but into the surrounding ecosystem: more reliable, easier to understand, and with your own sources.