# How Deep Research Works Mukund Sridhar, Google DeepMind & Aarush Selvan, Google | AI Engineer Summit 2025 | 15:15 Source: https://www.youtube.com/watch?v=eJOjdjO45Sc Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/how-deep-research-works Published: 2025-03-26 Tags: agents, context-engineering, planning, rag ## TL;DR - Gemini Deep Research spends several minutes browsing the web and building a cited report for questions where a chatbot would otherwise provide only a general plan. - The system plans iteratively, deciding what it can research in parallel, checking partial findings, and resolving missing or conflicting information across sources. - The product needs asynchronous UX, failure recovery, context management, and user-specific research workflows as tasks become longer and more capable. ## Summary Aarush Selvan and Mukund Sridhar explain the product and engineering decisions behind Gemini Deep Research. The system targets questions that need a comprehensive answer, such as athletic scholarship requirements or nuclear reactor supply chains. It creates an editable plan, browses the web for several minutes, shows activity while it works, and produces a long report with sources that can be edited and queried. Mukund focuses on the underlying agent loop. The model must split work into parallel and sequential parts, use partial findings to choose its next step, combine facts from different pages, resolve entities, handle inconsistent web layouts, recover from service failures, and manage context across follow-up tasks. The speakers then describe future directions: deeper domain expertise, personalized research workflows, and the ability to combine web research with coding, data science, financial modeling, and video generation. ## Key ideas ### Deep Research trades waiting time for a more complete answer [01:10](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=70s) The speakers saw that difficult questions often produced a blueprint rather than an answer. For example, a question about earning an athletic scholarship in shot put might receive advice to contact coaches, check throwing distances, and keep grades high. A useful answer would include the actual grade boundaries and qualifying distances. Deep Research removes some inference-time limits so Gemini can browse as much as needed and spend up to about five minutes on the task. The product is aimed at research and learning questions where the extra time can produce a comprehensive report. ### The product makes a long-running task visible before and during execution [01:59](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=119s) Gemini is usually a synchronous chatbot, so a multi-minute research task needs a different interaction model. Deep Research first shows an editable research plan in a card. This tells the user that the task will take time and lets them change the direction before it starts. During execution, the product shows the websites Gemini is browsing. Users can open those pages while they wait. The speakers said some users even tried to push the displayed website count into the thousands. ### A report becomes more useful when users can work inside it [04:33](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=273s) The final output can run to thousands of words, so Deep Research uses an artifact-like presentation inspired by Anthropic's artifacts. Users can pin the report and ask questions about it without repeatedly scrolling between the report and the chat. They can ask for a different style, add or remove sections, and ask follow-up questions. The system also shows the sources it read and the sources it used in the report. Sources remain available for follow-up questions and can carry over as citations when the report is exported to Google Docs. ### Long-running agents need state and recovery across many service calls [05:29](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=329s) A research task can run for minutes, make many language-model calls, and call services with different reliability. Mukund argues that a single intermediate failure should not discard the whole task. The system therefore needs state management and a way to recover from errors. The same design supports work that continues after the user leaves. Users could start a task, move to another device, and receive a notification when the research is ready. Mukund also considers future tasks that run for hours. ### The planner uses partial findings to choose the next research step [07:17](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=437s) The model must decide which parts of a question can run in parallel and which depend on earlier results. It also has to inspect what it has already found before planning the next action. In the shot-put example, finding Division I qualifying standards still leaves Division II and Division III unresolved. In a roller-coaster search, a page listing popular rides may say nothing about whether they suit children, so the planner needs another step to resolve that gap. Research is therefore an iterative loop rather than one search followed by one answer. ### Useful research often requires joining facts from separate pages [08:59](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=539s) Information for one answer is frequently spread across different sources. In the scuba certification example, one source explains the certification process while another gives the dive center's prices. The model must combine those details to work out the cost structure. It also faces entity resolution when different pages mention what may be the same organization or event. The system needs to use available indicators and investigate further when identity is uncertain. ### Browsing has to handle fragmented and inconsistent web pages [10:26](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=626s) Websites can describe the same subject with very different layouts. The speakers compare two pages about music festivals in Portugal. One presents most of the needed information in one place, while the other uses a different structure. A research agent therefore needs browsing mechanisms that can locate and interpret information across varied page designs. This problem affects the whole research process because the planner depends on the information each page exposes. ### Context management becomes harder with follow-up research [10:49](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=649s) Every search and intermediate result adds information to the model's context. Follow-up requests add more pressure because users may ask for the same kind of research on another topic. Mukund describes a strategy that keeps more detail about the current and recent tasks while selectively extracting older material into research notes stored in a retrieval system. This gives the model access to older work without carrying every detail forward at full size. Even with Gemini's long-context models, the system still needs explicit choices about what to retain. ### Future research agents must adapt their work to the user and domain [12:11](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=731s) The speakers describe Deep Research as a text-in, text-out system that retrieves from the open web, then outline several expansions. A more advanced agent would reason about implications and recommendations like a specialist partner, rather than only aggregate and synthesize information. It would also adapt the research and presentation to the user. A company due-diligence report for an individual might focus on strategic position, while a banker might need financial details and a discounted cash flow model. Future agents could combine browsing with coding, data science, financial modeling, and video generation. ## Notable quotes - "What if you remove the constraints of compute and latency at inference time, let Gemini take as long as it wants, browse the web as much as it needs, and see if we can trade that off for a much comprehensive answer for the user." (01:59) - "The first thing the model has to do is try to figure out which of these sub problems it can start tackling in parallel versus things that are inherently sequential." (07:37) - "So this notion of being able to ground on information you find and then plan your next step is key." (08:18) - "Just because you build something that can be really smart doesn't mean that it's useful to someone." (13:23) - "The way you browse the web, the way you frame your answer, the kind of questions you pursue should be very personalized to kind of meeting the user where they're at." (13:57) ## Tools & references mentioned - Gemini Deep Research - Gemini - Google DeepMind - Google Docs - Anthropic - Artifacts - Google Assistant - Goldman Sachs - McKinsey ## Who should watch - You are building an agent that runs for minutes or hours and need a plan for state, retries, and recovery between model and service calls. - Your application has to browse messy web pages, combine facts from separate sources, or resolve whether different pages refer to the same entity. - You are designing a research product and need concrete ideas for planning previews, progress visibility, long reports, citations, and follow-up questions. ## Related talks - [Enterprise Deep Research: The Next Killer App for Enterprise AI](https://aietalks.com/talks/enterprise-deep-research-the-next-killer-app-for-enterprise-ai) (Ofer Mendelevitch, Vectara, 05:19) - [How Google DeepMind Is Researching the Next Frontier of AI for Gemini](https://aietalks.com/talks/how-google-deepmind-is-researching-the-next-frontier-of-ai-for-gemini) (Raia Hadsell, Google DeepMind, 20:37) - [A year of Gemini progress + what comes next](https://aietalks.com/talks/a-year-of-gemini-progress-what-comes-next) (Logan Kilpatrick, Google DeepMind, 11:58) - [Building Conversational Agents](https://aietalks.com/talks/building-conversational-agents) (Thor Schaeff & Philipp Schmid, Google DeepMind, 1:47:34) - [Building in the Gemini Era](https://aietalks.com/talks/building-in-the-gemini-era) (Kat Kampf & Ammaar Reshi, Google, 17:57)