AI Engineer Summit 2025
How Deep Research Works
Loading slides…
Read the talk
How Deep Research Works
A shot-put scholarship question exposes what a research agent must do: plan around missing evidence, survive failures, combine sources, and preserve context for follow-up work.
From a talk by Mukund Sridhar and Aarush Selvan
Before you start: Basic familiarity with LLM prompts, context windows, and web retrieval will help with the architecture discussion.
From scholarship advice to concrete requirements
What does it take to get an athletic scholarship for shot put, and how do you get one? Answering that question requires more than advice about where to look. Gemini Deep Research was built to browse the web and produce a report on the user’s behalf. In the version presented by Aarush Selvan and Mukund Sridhar, users selected 1.5 Pro with Deep Research in Gemini Advanced. Selvan’s rounded $20 access price refers to the historical subscription, not a charge for each report.
Research and learning were already leading uses of Gemini, but difficult questions often produced a blueprint for finding an answer. For the shot-put question, that meant advice to contact coaches, find out how far to throw, and maintain good grades. The missing information was precisely what the user needed: actual academic requirements and throwing standards. A useful research agent must resolve the details that ordinary advice leaves as homework.
The engineering opportunity was to spend more inference time and make more browsing calls in exchange for a more comprehensive answer. That relaxation still had a practical budget: Selvan joked that the agent had to finish in five minutes because they did not have the chips for more. This was a framing of the compute constraint, not a universal completion-time guarantee.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An editable plan makes the wait intentional
A chatbot normally feels synchronous: ask a question, then read its answer. A research task that runs for minutes breaks that expectation. It also needs to be chosen deliberately. Waiting several minutes for the weather or a joke offers little benefit, while a long research report creates a second problem: how to read and work with thousands of words inside a conversation.
The product walkthrough uses a venture-capital research question: investigate recent breakthroughs in small nuclear reactors and identify interesting companies in their supply chain. Instead of immediately browsing, Gemini presents a research-plan card. The card makes the proposed work visible and gives the user a chance to change its scope.
The interaction proceeds through an explicit handoff:
- Review the approach. The plan describes the research Gemini proposes to undertake.
- Edit the plan. The user can steer the investigation before it begins.
- Start research. Approval moves the task from a proposal into execution.
Like an analyst explaining an approach before beginning, the agent establishes what it intends to do. The extra interaction also signals that this is a different kind of request from an ordinary chat response.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make progress inspectable and the report persistent
After approval, Gemini shows the websites it is browsing in real time. This interface predates thinking-model displays; its transparency comes from exposing source activity. Users can open those pages and read while the research continues. The count also created an unintended incentive: Selvan reports that some users deliberately pushed the displayed website count into the thousands. That anecdote describes people testing the counter, not a typical research workload or a measure of answer quality.
When research finishes, the report stays pinned beside the conversation. Selvan explicitly credits Anthropic’s Artifacts interface as inspiration. Keeping the document in view lets the user read a section and ask about it without scrolling back and forth through the chat. The same arrangement supports changes to writing style, adding or removing sections, and follow-up questions.
Sources consulted and sources cited are different sets. The interface exposes both, supporting user trust and attribution to publishers. A page can be read without contributing to the final report, yet its information can remain in context for later questions. Citations also carry into Google Docs when the user exports the report, preserving the connection between the document and its supporting sources.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The task must outlive an individual call
Behind the interface are four connected engineering concerns: long-running work, iterative planning, interaction with a noisy web, and growing context. Each affects the others. Browsing produces intermediate information; that information changes the plan; the revised plan triggers more calls and expands the task’s state.
A research task can make many LLM and service calls over several minutes. Intermediate failures are therefore part of the operating environment. Sridhar also anticipates tasks that could eventually run for hours, making it even more important to recover from a failed service without dropping the whole investigation. State management must preserve the research task across failures of its individual operations. The talk identifies that requirement without prescribing a particular storage system or retry implementation.
The same separation between a task and its immediate interaction enables a different usage pattern: submit the request, walk away, receive a notification, and read the result later. The user may return on another device. Recoverable task state therefore serves both reliability and the cross-device product experience.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Choose the next action from what is still missing
Return to the shot-put scholarship question. Its subproblems do not all have the same dependencies: some can be investigated in parallel, while others require earlier findings. The model must identify that structure, then inspect the information already collected before deciding what to do next. Parallel execution saves time only when the work is actually independent.
In Sridhar’s example, the agent has found Division I qualifying information but still needs to investigate Divisions II and III. The point is the gap in coverage, not that all divisions offer equivalent scholarships. Current NCAA guidance states that Division III schools do not offer athletics scholarships, although other aid may be available. A complete investigation must distinguish participation requirements from scholarship availability instead of filling every division with the same kind of answer.
A small TypeScript representation makes this planning boundary concrete. The records below track evidence coverage, not athletic eligibility; the resulting actions are proposed searches, not completed findings.
typescript
type Division = "D1" | "D2" | "D3";
type Coverage = {
division: Division;
qualifyingInformation: "found" | "unresolved";
};
const coverage: Coverage[] = [
{ division: "D1", qualifyingInformation: "found" },
{ division: "D2", qualifyingInformation: "unresolved" },
{ division: "D3", qualifyingInformation: "unresolved" },
];
const proposedActions = coverage
.filter(item => item.qualifyingInformation === "unresolved")
.map(item => ({
division: item.division,
status: "proposed" as const,
question:
`For ${item.division}, verify shot-put participation requirements ` +
"and whether athletics scholarships are available.",
}));
The general loop is to ground the next plan in accumulated evidence. Finding one answer changes what remains worth investigating.
A second example exposes a subtler gap. A search for the best roller coaster for children returns a page listing the top ten roller coasters. The page is relevant to roller coasters, but that alone does not establish suitability for children. The planner must recognize the unresolved constraint and pursue it in later steps. Topical relevance is not the same as answering the question.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Combine facts without confusing their sources
Even when the necessary facts exist, they may not appear together. For a question about scuba certification at nearby dive centers, one source explains the course structure and another gives the center’s pricing. The agent has to combine those facts to determine the certification’s cost structure. Neither page independently supplies the whole answer: the useful result comes from relating the requirements to the relevant prices.
Combining sources also introduces entity resolution. Two pages may mention what appears to be the same organization or offering, but the agent needs identifying information to determine whether those mentions really refer to one entity. When the evidence remains ambiguous, further research is part of the task. Otherwise, a plausible synthesis can attach one entity’s facts to another.
Accessing the evidence is a separate challenge from interpreting it. Sridhar compares two websites about music festivals in Portugal: one makes most of the information available in a single view, while the other uses a different layout. A robust browsing mechanism must navigate these variations. Search can locate a relevant page, but the agent still needs to reach and read the information within that page’s structure.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep recent detail and retrieve older research
Each research step adds intermediate outputs to the context. Finishing the first report does not end that growth: a user may ask a follow-up question or request the same investigation for another topic. Those subsequent tasks add their own evidence and planning history. Gemini’s long-context models provide room, but they do not remove the need to decide what stays available directly.
Sridhar presents one approach with a deliberate recency bias:
| Research history | Treatment |
|---|---|
| Current task | Keep richer information in context |
| Previous task | Retain more detail |
| Older tasks | Select research notes for retrieval |
The older notes go into a retrieval-augmented generation (RAG) system, so the model can still access them selectively. This reduces how much old material must remain in active context without making the older work entirely inaccessible. It is one context-management choice among several, with its own trade-offs.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Research that changes with the user and the task
At the December launch, the team was unsure whether users would accept waiting several minutes for an answer. Selvan reports a positive reception and compares the product to a McKinsey analyst available through an inexpensive subscription. That is his analogy for the product’s value, not a validated performance comparison. The system he describes here retrieves from the open web and accepts and produces text; those boundaries belong to the talk-era product.
The first proposed direction is deeper expertise. Moving from an analyst to a partner, in Selvan’s analogy, means going beyond aggregation and synthesis to explain implications: which patterns matter, what they suggest, and how they affect a decision. In science, the corresponding ambition is to read many papers, compare methods, recognize patterns, and propose new hypotheses to investigate. These are directions for development rather than capabilities demonstrated in the walkthrough.
Expertise alone does not make an answer useful to a particular person. Consider company due diligence:
| Reader | Information needed |
|---|---|
| General reader | Company explanation and strategic position |
| Banker | Detailed financials and a discounted cash-flow model |
A banker’s needs imply more than a different writing style. They require finer-grained financial analysis and should change which questions the agent pursues and which sources it browses. Personalization belongs in the investigation itself, not only in the final presentation.
The final direction is to combine web research with coding, data science, and even video generation. In the due-diligence example, an agent could perform statistical analysis and construct financial models, then use those results to inform the research output. The report would draw on work performed over the evidence, as well as on retrieved text. Selvan explicitly qualifies the financial example: Google does not give financial advice and is not a financial advisor. The proposed expansion is toward research whose tools and outputs fit the work the user actually needs to do.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Google’s original description of Deep Research, including editable plans, iterative web research, linked reports and Google Docs export.
Further reading
Anthropic’s June 2024 explanation of working with generated documents and code in a separate panel alongside a conversation.
- Division III eligibility and financial aidDocumentation
Official NCAA guidance distinguishing Division III participation and academic requirements from athletics scholarships.
Full transcript
Loading transcript…