Google Launches Upgraded Deep Research Agent Powered by Gemini 3 Pro
This release hints at how research and information-gathering may soon work behind the scenes.
Google clearly didn’t want to watch the spotlight drift to OpenAI this week. While the industry held its breath for the launch of GPT-5.2 “Garlic,” Google pushed out something with far bigger long-term implications: a rebuilt version of its Gemini Deep Research agent.
On the surface, it looks like a research assistant. In reality, it signals Google’s attempt to redefine how information is gathered, verified, and automated in the coming age of agentic AI.
Emmanuel Umahi
The story starts with Gemini 3 Pro, Google’s newest reasoning model. Deep Research is now powered directly by this engine, meaning it can read huge piles of documents, connect them, and spit out structured insights with far fewer factual slip-ups than previous generations. Google keeps describing it as “factual,” but the real shift is deeper: they want AI that can stay accurate across tasks that run for minutes or even hours without falling apart midway.
To make that possible, Google built a new Interactions API that allows developers to pull this research capability straight into their own tools. Instead of single prompts and single answers, the API lets AI agents hold long sessions, manage context on the server side, plan multiple steps, and recover when the task becomes messy. Developers can feed it PDFs, datasets, website links, and the agent figures out what it needs, what’s missing, and where to dig next.
This is Google responding to a problem that everyone in AI research already knows. Large models can talk, but they often fold under pressure during extended reasoning. A single wrong assumption early on can tank an entire report. Google’s upgrade tries to fix that by making each step transparent, sourcing every claim, and allowing the agent to re-check its own work. It’s the closest Google has come to building an AI that behaves like a patient researcher rather than a fast typist.
To prove the improvements, Google leaned heavily on benchmarks. It introduced a new one, DeepSearchQA, and open-sourced it for the world to poke holes in. On Humanity’s Last Exam, the notoriously punishing benchmark of obscure general knowledge, Deep Research reached 46.4%. On Google’s own new DeepSearchQA, designed for multi-step research tasks, it posted 66.1%. And on BrowseComp, the benchmark for web-based agent tasks, it landed 59.2%.
But the numbers were barely out before OpenAI’s Garlic arrived, instantly challenging Google’s claims and outperforming Gemini on several existing metrics. That back-and-forth has become the rhythm of the industry: every benchmark win lasts only until the next press release.
Where this announcement becomes interesting is how quickly Google plans to weave Deep Research into its ecosystem. Soon, the same agent devs are experimenting with will start assisting Google Search, Google Finance, NotebookLM, and even the Gemini app. The direction is clear: humans will gradually outsource information-gathering to AI agents, and Google wants to be the infrastructure behind that shift.
The broader industry is inching toward a future where the search box feels outdated. Instead of typing queries, users might delegate entire goals, “analyse this company,” “summarise this dataset,” “map the risks.” Google’s latest update isn’t the flashy moment that signals this future, but it’s one of the building blocks that will quietly make it possible.
David Adubiina
