Google Releases Upgraded Deep Research Agent With New Interactions API For Developers

In Brief

Google just released an upgraded version of its Deep Research agent, now available to developers through a new Interactions API — with consumer rollouts coming soon to Search, NotebookLM, and the Gemini app.

Technology company Google stated that it has released a substantially upgraded version of its Deep Research agent, now accessible to developers through a new Interactions API, with consumer availability planned for Search, NotebookLM, and the Gemini application.

For the first time, developers are able to integrate Google’s most advanced autonomous research capabilities directly into their own applications. Gemini Deep Research is designed for extended information-gathering and synthesis tasks, and its reasoning system is powered by Gemini 3 Pro, described as the company’s most factual model to date. It has been trained to reduce hallucinations and enhance the clarity and reliability of complex reports. By expanding multi-step reinforcement learning for search, the agent can independently navigate intricate information environments with improved accuracy.

The agent constructs its research workflow step by step by generating queries, reviewing results, identifying missing information, and continuing the process until it completes its investigation. The new release includes major upgrades to web search performance, enabling deeper navigation into websites to extract highly specific data.

According to Google, the latest version delivers state-of-the-art performance on Humanity’s Last Exam (HLE) and DeepSearchQA, while also achieving its strongest results to date on BrowseComp. It is optimized for producing well-researched reports at significantly lower cost and will soon be integrated into Google Search, NotebookLM, Google Finance, and an enhanced version of the Gemini application.

Early testing already shows substantial gains across fields where accuracy and detailed contextual understanding are essential. In financial services, firms have begun using Gemini Deep Research to streamline the early phases of due diligence by aggregating market indicators, competitor insights, and compliance considerations from both public and proprietary sources. This has made the agent a valuable tool for investment teams conducting preliminary workflows.

Within the scientific sector, the agent is being applied to complex safety-related research. Axiom Bio, a company developing AI systems for predicting drug toxicity, reported that Gemini Deep Research provided a depth of initial analysis and precision across biomedical literature that allowed its research and discovery processes to progress more rapidly.

For developers building automated research systems, the Gemini Deep Research agent offers broad functionality for synthesizing information and producing detailed, verifiable reports. It supports unified analysis of user documents such as PDFs, CSVs, and text files alongside public web sources by combining File Upload with the File Search Tool.

It manages extensive context effectively, enabling developers to include large amounts of background material directly in the prompt. Output structure can be shaped through prompting, allowing full control over report layout, headings, and data presentation. The system provides granular citations for claims, ensuring transparency regarding data provenance, and supports structured outputs, including JSON schemas, for streamlined integration into downstream applications.

Google Open-Sources DeepSearchQA Benchmark To Advance Multi-Step Web Research Capabilities

Additionally, Google announced the open-sourcing of a new benchmark called DeepSearchQA, created to evaluate how effectively research agents handle comprehensive, multi-step web-based inquiry. DeepSearchQA includes 900 manually constructed causal-chain tasks spanning 17 subject areas, with each step building on the conclusions of the previous one. Rather than relying on simple fact-retrieval questions, the benchmark measures an agent’s ability to produce complete and exhaustive answer sets, enabling assessment of both research accuracy and retrieval coverage.

DeepSearchQA is also intended as a diagnostic resource for studying the effects of extended reasoning time. Internal testing has shown that performance improves when agents are given more opportunities to run additional searches and reasoning cycles, an area Google expects to expand on in future iterations.

The benchmark materials are being released to encourage continued progress toward more capable research agents. Developers and researchers can review the dataset, leaderboard, and starter Colab, as well as examine the underlying methodology described in the accompanying technical report.

Although the Deep Research landscape is already highly competitive, Google’s updated agent introduces notable enhancements that build on the capabilities of the existing Gemini 3 models. The release also marks the first time developers can integrate this technology directly into their own applications, offering a significant improvement to the research functionality within third-party products.