Gemini’s Upgraded Deep Research & Agent Now Available
▼ Summary
– Google has announced a new, more powerful Gemini Deep Research agent, now available for developers and coming soon to consumer apps like Gemini and Google Search.
– A new Interactions API provides a unified interface for developers to connect Gemini models, Google’s built-in agents, and their own custom agents.
– The Gemini Deep Research agent is optimized for complex, long-running research tasks, using Gemini 3 Pro to reduce errors and improve report quality.
– This agent autonomously performs multi-step web research by formulating queries, reading results, identifying knowledge gaps, and conducting deeper searches.
– It demonstrates state-of-the-art performance on key benchmarks for reasoning, web research, and fact-finding, surpassing the base Gemini 3 Pro model.
Google has unveiled a significantly enhanced version of its Gemini Deep Research agent, marking a major step forward for developers and a preview of what’s to come for everyday users. This upgraded tool, powered by the advanced Gemini 3 Pro model, is designed to tackle complex, long-form research tasks with greater autonomy and accuracy. The announcement centers on a new Interactions API, which provides a single, unified gateway for developers to connect with Google’s models and a growing suite of intelligent agents.
The core of this release is the Gemini Deep Research agent, now available in preview for third-party developers to integrate into their own applications. This agent is optimized for gathering and synthesizing information from extended contexts, a process that involves formulating queries, reading results, identifying knowledge gaps, and conducting follow-up searches autonomously. A key improvement is its vastly enhanced web search capability, enabling it to delve deeply into websites to locate specific, hard-to-find data points. Google emphasizes that the agent has been specifically trained to minimize incorrect information, or hallucinations, and to maximize the quality of its final reports during these intricate tasks.
This performance is achieved by scaling multi-step reinforcement learning for search, allowing the agent to navigate complex information landscapes with high precision. The results are demonstrated on several demanding benchmarks. On Humanity’s Last Exam, which tests reasoning and broad knowledge, the Deep Research agent scored 46.4%, outperforming the base Gemini 3 Pro model’s 43.2%. Its prowess in comprehensive web research is shown on the DeepSearchQA benchmark, where it achieved 66.1% versus 56.6%. Perhaps most impressively, for the challenging task of locating obscure facts in BrowseComp, the agent reached 59.2%, a substantial leap from the 49.4% score of Gemini 3 Pro.
For developers, the new framework also introduces the ability to build and incorporate custom agents alongside Google’s built-in offerings. The Interactions API is the conduit for this, enabling a seamless connection between Gemini models, Google’s pre-built agents like Deep Research, and an organization’s own specialized agents. This creates a flexible ecosystem for developing sophisticated AI-driven applications.
While developers can begin experimenting with these capabilities immediately in Google AI Studio, these powerful enhancements are not staying in the developer sandbox for long. Google has confirmed that the upgraded research and agent functionalities will “soon” make their way into popular consumer applications. This means users can expect to see these advanced autonomous research features integrated into platforms like Gemini, Google Search, and NotebookLM, bringing a new level of AI-assisted discovery and synthesis to a wider audience.
(Source: 9to5 Google)
