# How Deep Research Works

Mukund Sridhar, Google DeepMind · Aarush Selvan

AI Engineer Summit 2025

[Talk page](https://ai.engineer/talks/eJOjdjO45Sc-deep-research-works) · [Video](https://www.youtube.com/watch?v=eJOjdjO45Sc)

## 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.

Basic familiarity with LLM prompts, context windows, and web retrieval will help with the architecture discussion.

### From scholarship advice to concrete requirements

[0:34](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=34s)

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](https://blog.google/products-and-platforms/products/gemini/google-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.

![Slide titled “What is Deep Research” showing Gemini’s model menu with “1.5 Pro with Deep Research” selected and a panel reading “Researching 56 websites...”](https://ai.engineer/api/talk-essay-media/e517e3ad943796a1945810877224eade4983106d15dd9b1b128b5717972320a4)

Deep Research in Gemini’s model menu, alongside a research progress display.

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.

### An editable plan makes the wait intentional

[2:18](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=138s)

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:

1. **Review the approach.** The plan describes the research Gemini proposes to undertake.
2. **Edit the plan.** The user can steer the investigation before it begins.
3. **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.

![Gemini displays a “Small Nuclear Reactor Advancements” plan with Research Websites, Analyze Results, Create Report, and “Edit plan” and “Start research” buttons.](https://ai.engineer/api/talk-essay-media/632e727d2d9078456c1d542c89a730f4a704810c0e8c5e14209d5f8b3c35eaee)

A small nuclear reactor research plan with controls to edit the plan or start research.

### Make progress inspectable and the report persistent

[3:51](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=231s)

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](https://www.anthropic.com/news/projects) 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.

### The task must outlive an individual call

[5:29](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=329s)

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.

### Choose the next action from what is still missing

[7:21](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=441s)

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](https://www.ncaa.org/eligibility-center/initial-eligibility-requirements/division-iii/) 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.**

### Combine facts without confusing their sources

[9:05](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=545s)

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.

![“Challenges: Multi Step Planning” slide with a scuba certification prompt, course overview and class pricing screenshots, and two blue arrows pointing to “Calculate course cost.”](https://ai.engineer/api/talk-essay-media/4b21ee561284d9deb00cfafa09563eba0c8348307bd26ae478bef20329d4e79b)

Course requirements and pricing from separate pages feed into a course-cost calculation.

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.

### Keep recent detail and retrieve older research

[10:44](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=644s)

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.

### Research that changes with the user and the task

[12:00](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=720s)

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.

![“What’s Next” slide contrasts open-web retrieval and text input/output with greater expertise, personalized research, and combining web research with coding, data science, and video generation.](https://ai.engineer/api/talk-essay-media/dc1eb06f79fdfc423d0f7bfeefeeaeb8c2d040c7e9e05f23ecd02c40855ed8ce)

The research-agent roadmap adds expertise, personalization, and work across domains.

## Resources

- [Gemini Deep Research launch announcement](https://blog.google/products-and-platforms/products/gemini/google-gemini-deep-research/) — Google’s original description of Deep Research, including editable plans, iterative web research, linked reports and Google Docs export.
- [Claude Projects and the Artifacts interface](https://www.anthropic.com/news/projects) — Anthropic’s June 2024 explanation of working with generated documents and code in a separate panel alongside a conversation.
- [Division III eligibility and financial aid](https://www.ncaa.org/eligibility-center/initial-eligibility-requirements/division-iii/) — Official NCAA guidance distinguishing Division III participation and academic requirements from athletics scholarships.

## Slides

### What is Deep Research

Gemini’s Deep Research browses the web to build reports. The slide illustrates a research run across 56 websites.

[0:31](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=31s) [Gemini Advanced model menu alongside an illustration of Deep Research researching websites.](https://assets.ai.engineer/sites/aiecode2025/sha256/2f02371808c689e8f742b097e1a8a32fa7a5b920ac835e52aed779f53bfc313f.webp)

What is Deep Research

Gemini Advanced
1.5 Pro with Deep Research

Gemini Advanced

2.0 Flash
For everyday tasks, plus more features

2.0 Flash Thinking Experimental
Best for multi-step reasoning

2.0 Flash Thinking Experimental with apps
Reasoning across YouTube, Maps & Search

2.0 Pro Experimental
Best for complex tasks

1.5 Pro with Deep Research
Get in-depth answers

Researching 56 websites...

Sequences shortened and screen images simulated.

Confidential - Google DeepMind

### Product Challenges in Building a Research Agent

Deep Research brought asynchronous work to a synchronous chatbot, required clearer expectations for everyday queries, and could produce outputs of 3k+ words that are difficult to navigate.

[2:18](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=138s) [Product challenges include asynchronous interaction, setting user expectations, and handling long outputs.](https://assets.ai.engineer/sites/aiecode2025/sha256/8da908d79c45e11ce02f93ba6172724c07f84a24b69ff2aeea9d2352a0cf91ef.webp)

Product Challenges in Building a Research Agent

Gemini’s first asynchronous feature
The app is inherently designed around a synchronous product experience

Setting expectations with users
Deep Research is not helpful for most of the common queries that users come to Gemini for

Handling long outputs
Outputs can be 3k+ words, which can be unwieldy to engage with

Interface text:
Confidential - Google DeepMind

[2:35](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=155s) [Slide lists challenges involving asynchronous research, user expectations, and long outputs.](https://assets.ai.engineer/sites/aiecode2025/sha256/00185fc27a633c3ab53f4185b1740f6fde50b8673c06fc6150278582f6b9cdf9.webp)

Product Challenges in Building a Research Agent

Gemini’s first asynchronous feature
The app is inherently designed around a synchronous product experience

Setting expectations with users
Deep Research is not helpful for most of the common queries that users come to Gemini for

Handling long outputs
Outputs can be 3k+ words, which can be unwieldy to engage with

Interface text:
Confidential - Google DeepMind
AIE
AGENT ENGINEERING
HTTPS://AI.ENGINEER

### Example Research Query

A hypothetical VC asks Deep Research about the latest technology breakthroughs in small nuclear reactors and the most interesting companies in their supply chain.

[3:08](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=188s) [A prompt asks about small nuclear reactor technology and companies in its supply chain.](https://assets.ai.engineer/sites/aiecode2025/sha256/3e812bc5e6848ff933542815deeedd9a8faf5b79e7659e89eb1e0dd182e2ff71.webp)

Help me learn about the latest technology breakthroughs in small nuclear reactors and the most interesting companies in the supply chain of this industry

Interface text:
Confidential - Google DeepMind
AIE
AGENT ENGINEERING
HTTPS://AI.ENGINEER

### Small Nuclear Reactor Advancements

Gemini presents an editable research plan before starting. Once research begins, it shows the websites being browsed; the demo displays 46 in progress.

[3:19](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=199s) [Gemini displays a research plan for small nuclear reactor advancements, with options to edit the plan or start research.](https://assets.ai.engineer/sites/aiecode2025/sha256/5f09fe07ffb9914db892390e8fec41dbd29297ecbd744c9f082589e2f6c10fdb.webp)

Here's a research plan for that topic. If you need to update it, let me know!

Small Nuclear Reactor Advancements

Research Websites
Help me learn about the latest technology breakthroughs in small nuclear reactors and the most interesting companies in the supply chain of this industry by:
(1) Find articles and research papers on the latest technology breakthroughs in small nuclear reactors.
(2) Find information on the current state of the small nuclear reactor industry.
(3) Find information on the supply chain of the small nuclear reactor industry....
More

Analyze Results
Create Report
Ready in a few mins

Interface text:
Confidential - Google DeepMind
Gemini Advanced
1.5 Pro with Deep Research
Edit plan
Start research
Ask Gemini
Gemini can make mistakes, so double-check it

[3:53](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=233s) [Gemini shows a research task in progress alongside a list of 46 websites.](https://assets.ai.engineer/sites/aiecode2025/sha256/e344b5d2860e3e7a9efb517df4a9d93f27d1f7a8acee24d0faaf03a88fb85462.webp)

Start research

Great. While I'm researching, feel free to leave this chat. I'll let you know as soon as I'm done.

Small Nuclear Reactor Advancements
Researching 46 websites...

Researching 46 websites...
In researching your report, Gemini searches the web and browses content from relevant websites on your behalf.

Top 10 Startups Revolutionizing Data Center Energy ... — bisresearch.com/insights/Top-10-startups-revolutioni...
Top 9 Nuclear Power Companies Leading the Global... — energy-oil-gas.com/news/top-10-nuclear-power-com...
Top Nuclear Energy Companies Shaping the Future ... — vaneck.com/us/en/blogs/natural-resources/top-nucle...
Small Nuclear Power Reactors - World Nuclear Asso... — world-nuclear.org/information-library/nuclear-fuel-cy...
Five of the World's Leading Small Modular Reactor ... — c3newsmag.com/five-of-the-worlds-leading-small-m...
Nuclear Small Modular Reactors (SMRs) 2023-2043 ... — idtechex.com/en/research-report/nuclear-small-mod...
SMR Technology Trends Worldwide - Enerdata — enerdata.net/publications/executive-briefing/smr-world-t...
Global small modular reactor pipeline hits 22 GW, wi... — utilitydive.com/news/small-modular-reactor-haleu-ge...
Small modular reactors (SMR) | IAEA — iaea.org/topics/small-modular-reactors

Interface text:
Confidential - Google DeepMind
Gemini Advanced
1.5 Pro with Deep Research
Ask Gemini
Gemini can make mistakes, so double-check it

### Gemini Deep Research report and sources

Gemini pins the completed report beside the chat for follow-up questions. The report shows its sources and 43 researched websites; the speaker says not every page read is used in the report.

[4:32](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=272s) [Gemini displays a completed report on small nuclear reactors beside a chat for follow-up questions.](https://assets.ai.engineer/sites/aiecode2025/sha256/50d4d64478fbba73f388822e25b242d58f5145f347c473fbbeee4b6fd81eeef8.webp)

Small Nuclear Reactors: Technological Breakthroughs and Leading Companies

Introduction

The global energy landscape is undergoing a significant transformation, driven by the urgent need for clean and reliable energy sources. Small modular reactors (SMRs) are emerging as a key player in this transition, offering a compelling solution to meet the world's growing energy demands while minimizing carbon emissions. SMRs are advanced nuclear reactors that are significantly smaller than conventional nuclear power plants, typically generating up to 300 megawatts of electricity (MWe) per unit. These reactors are designed with modularity in mind, allowing for factory fabrication and assembly, which reduces construction costs and timelines. This article explores the latest technology breakthroughs in SMRs, the current state of the industry, the supply chain, and highlights some of the most interesting companies involved.

Technology Breakthroughs in Small Nuclear Reactors

Advanced Reactor Designs

The field of SMRs is witnessing a surge in innovative reactor designs, each with unique features and benefits. Some of the most notable advancements include:

Interface text:
Gemini Advanced
1.5 Pro with Deep Research
Start research
I've completed your research. Feel free to ask me follow-up questions or request changes.
Small Nuclear Reactor Advancements
Feb 20, 11:05 PM
Small Nuclear Reactor Advancements
Open in Docs
Ask Gemini
Gemini can make mistakes, so double-check it

[5:06](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=306s) [Research report showing source cards, a list of researched websites, and a citation example.](https://assets.ai.engineer/sites/aiecode2025/sha256/23b543f6d02361ec6def06bd53dc1d80a90e7c0291e4a360c4d08344874adb7f.webp)

enhance energy security, and drive economic growth. While the industry faces challenges related to licensing, supply chain, waste management, and public perception, ongoing technological advancements, government support, and industry collaboration are paving the way for the successful deployment of SMRs. As the global energy landscape continues to evolve, SMRs are poised to play a crucial role in meeting the world's growing energy demands while minimizing carbon emissions.

Sources and related content

Advanced Small Modular Reactors (SMRs) - Department of Energy
e​nergy.gov/ne/advanced-s...

The small reactor revolution can transform African energy systems...
atlanticcouncil.org/content...

Small Nuclear Power Reactors - World Nuclear Association
world-nuclear.org/informat...

Researched 43 websites

In researching your report, Gemini searches the web and browses content from relevant websites on your behalf.

Accelerating Advanced Nuclear Deployment: Meetin...  iaea.org/newscenter/news/accelerating-advanced-nu...
Supply Chain Map - Nuclear Reactor Components  nei.org/CorporateSite/media/filefolder/Supply-Chain-Map...
Management of the nuclear supply chain | IAEA  iaea.org/topics/management-systems/management-of-th...

• Passive Cooling Systems: Many SMR designs utilize passive cooling systems [illegible]
like gravity and convection to remove heat from the reactor core, minimizing [illegible]
Passive cooling systems eliminate the need for active components like pump[illegible]
for failures and increasing reliability. These systems rely on natural phenome[illegible]
where heated fluids rise and cooler fluids sink, to circulate the coolant and re[illegible]
reactor core.

1. The small reactor revolution can transform African energy...
atlanticcouncil.org/content...

• Below-Grade Placement: SMRs are often placed below ground level, providi[illegible]

Interface text:
Confidential - Google DeepMind
Small Nuclear Reactor Advancements
Open in Docs

### Engineering challenges in building a research agent

A research agent must survive long-running jobs, plan as it finds information, navigate a fragmented web, and manage growing context while balancing exploration against its inference budget.

[5:38](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=338s) [Four challenges: long-running tasks, multi-step planning, fragmented web content, and context management.](https://assets.ai.engineer/sites/aiecode2025/sha256/b913e9a2c4d2888568c644e458258cb3dd11aaa1bb9b03de54937cb658820b94.webp)

Eng Challenges in Building a Research Agent

Long Running Tasks
Need to build asynchronous system. Be resilient to failures of intermediate steps during long running tasks.

Multi Step Planning
Model needs to react to information found along the way and decide what needs further exploration / verification. Exploration vs inference budget trade off.

Fragmented web
Websites are noisy, common artifacts are encoded in variety of ways across websites, access issues, partial information

Context Management
Balancing between reasoning jointly over all information vs efficiency. Reusing research across turns when relevant.

Interface text:
Confidential - Google DeepMind

### Challenges: Long-running tasks

A research task can make many service calls over several minutes. State management helps it recover from an intermediate failure, while a notification tells the user when the report is ready.

[6:12](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=372s) [A sequence of service calls illustrates an intermittent failure, alongside an example report-ready notification.](https://assets.ai.engineer/sites/aiecode2025/sha256/b8af24286a70f1c49faad7514e60ee77f4e9c6e4550c1de6ec0b210300e242d2.webp)

Challenges : Long Running Tasks

Given jobs are O (minutes) with many calls to the model/services, we need a system that

Is robust to intermittent failures or periods of peak usage

RPC 36 → RPC 37 → RPC 38 😵

Notifies the user when the task is done

3:59
Thu, Feb 6  ⛅ 24°C
Report ready · Gemini · now
Small Nuclear Reactor Technology and In...

Interface text:
Confidential - Google DeepMind
AIE
AGENT ENGINEERING
HTTPS://AI.ENGINEER

### Multi-step planning: athletic scholarships

The model tracks completed scholarship research, uses partial information to search for other divisions’ standards, and asks a web page specific recruiting questions.

[7:21](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=441s) [A scholarship research example shows the model tracking completed tasks and identifying further searches.](https://assets.ai.engineer/sites/aiecode2025/sha256/80013ab5ea426ea3c96fb58e16a4746c1c8d0addbb62420432b147e7ff9c3850.webp)

Challenges: Multi Step Planning

My son is interested in going to college on an athletic scholarship for shot put. Help me make a plan to navigate this process

Model continuously reasons over which sub-tasks it has completed vs need more work

# I have completed the following steps:
# "(1) Find the top colleges with athletic scholarships for shot put", and
# "(3) Find the coaches for shot put at each college".

# Let's continue to work on the following step.
# "(2) Find the qualifying standards for shot put for each college"

# I have info that the qualifying standard for shot put is 66'3 for D1.
Let's find the standard for other divisions.
search: "D2, D3 college shot put recruiting standard"

Within each sub-task the model asks specific questions over relevant web pages it discovers

# Let's continue to work on the following step.
# "(5) Find the recruiting process for shot put at each college"

Browse: "tell me the recruiting process for shot put, deadlines for applying for athletic scholarships, financial aid options and academic requirements" from "https://throwslab.com/shot-put-scholarships/"

Confidential - Google DeepMind
Proprietary & Confidential

### Multi-step planning: roller coaster suitability

A top-10 coaster ranking identifies Iron Gwazi, but leaves its suitability for kids unanswered. The research plan must recognize that gap and investigate it in a later step.

[8:33](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=513s) [A roller coaster article is paired with a question about whether its featured ride is suitable for kids.](https://assets.ai.engineer/sites/aiecode2025/sha256/ac71b645a8f17ab2a117e1c9347cb4fdf526b0539a9f94c3d8d3850bd7ba4695.webp)

Challenges: Multi Step Planning

What are the best roller coasters in America for a family with kids?

Top 10 roller coasters in the USA
1. Iron Gwazi at Busch Gardens Tampa – RMC

Rocky Mountain Construction (RMC) has created a masterpiece at Busch Gardens Tampa and it is called Iron Gwazi. The intensity of pure thrill is combined with a constant flawless grace, stringing together element after element unlike any coaster on this planet. Iron Gwazi is a perfect roller coaster.

Highlights: The Death Spiral needs to be experienced to be believed. And the purple wash of lighting at night creates an unforgettable and dreamy ride.

Is it suitable for kids?

Interface text:
Confidential - Google DeepMind
blooloop.
Home  News  Directory  Blooloop 50  Innovation Awar

[8:57](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=537s) [A roller coaster ranking is paired with a question about whether the ride is suitable for children.](https://assets.ai.engineer/sites/aiecode2025/sha256/7b7e4a6675ce95e6bb927c9f9a9fdb93f3c26cafc2cfd29b6ed19c48a1bafc93.webp)

Challenges: Multi Step Planning

What are the best roller coasters in America for a family with kids?

Top 10 roller coasters in the USA
1. Iron Gwazi at Busch Gardens Tampa – RMC

Rocky Mountain Construction (RMC) has created a masterpiece at Busch Gardens Tampa and it is called Iron Gwazi. The intensity of pure thrill is combined with a constant flawless grace, stringing together element after element unlike any coaster on this planet. Iron Gwazi is a perfect roller coaster.

Highlights: The Death Spiral needs to be experienced to be believed. And the purple wash of lighting at night creates an unforgettable and dreamy ride.

Is it suitable for kids?

Interface text:
Confidential - Google DeepMind
blooloop.
Home  News  Directory  Blooloop 50  Innovation Awar
AIE
AGENT ENGINEERING
HTTPS://AI.ENGINEER

### Multi-step planning: scuba course cost

Estimating scuba certification costs requires combining the course structure on one page with pricing information on another.

[9:05](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=545s) [Course overview and pricing pages point toward calculating certification cost.](https://assets.ai.engineer/sites/aiecode2025/sha256/d98de1659172ff5c7294159834b9c64178678fe80c5c708ca97df1a6c6946c60.webp)

Challenges: Multi Step Planning

I'm interested in learning to scuba dive and want to get certified. Please research the best PADI Open”

Course Overview page

GET READY FOR YOUR OPEN WATER CERTIFICATION!
Are you looking to go diving underwater? We are here to help!
With Bamboo Reef Dive Centers you can complete your certification in 3 easy steps:

1. HOME STUDY
Complete your Open Water Diver Online eLearning from SDI at your leisure. With 6 dedicated chapter of in depth information you'll be pumped to start diving!

2. CLASS & POOL SESSIONS
Sign up for our Weekend or Evening OpenWater Classes. Over a period of 4 hours of class review and over 10 hours of pool time, you'll be set for the ocean!

3. OCEAN CERTIFICATION DIVES - MONTEREY
Join us down in beautiful Monterey, CA where you will complete 4 dives over 2 days and complete your Open Water journey.

OR
OCEAN CERTIFICATION DIVES - REFERRAL
We'll send you off with referral paperwork where you will finish your 4 required dives on vacation. Return your paperwork and we'll sign you off as an Open Water Diver.

Class pricing page

MAKEUPS/ADDITIONAL TRAINING:
If you miss a class or need additional time in the water Bamboo Reef and your instructor will work with you to setup a schedule.
Zoom/Classroom Sessions - $60.00 per hour
Pool Sessions - $80.00 per hour
MONTEREY: Pool Sessions do not include the lane reservations or entrance fees to the Monterey Sports Center
Ocean Dives - SCUBA
Rejoin a Group Session - $150.00 per day. A pool refresher is required if it has been more than 30 days since your last pool session.
Private 1-on-1 - $150.00 per dive
Ocean Dives - Freedive
Rejoin a Group Session - $75.00 per day. A pool refresher is required if it has been more than 30 days since your last pool session.
Private 1-on-1 - $150.00 per dive

MEDICAL WAIVERS:
Every student participating in a SCUBA/Freedive/Tech Diving class with Bamboo Reef is required to complete the RSTC Medical Waiver.
RSTC Medical waivers are good for 1 calendar year.

Calculate course cost

Interface text:
Confidential - Google DeepMind

### Challenges: Web is Fragmented

Two Portuguese music festival sites present information differently: one puts details on a single page, while the other requires one-hop browsing through listings.

[10:09](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=609s) [Two music festival websites illustrate direct information retrieval and browsing through a listing.](https://assets.ai.engineer/sites/aiecode2025/sha256/2336ec660238c7a19017c9024e2a92167578a18fc939e231f1abcd722cea82d9.webp)

Challenges: Web is Fragmented

Easy page to retrieve info

The best Portuguese music festivals in 2025

April 2025
Tremor Festival
When: 8–12 April 2025
Where: São Miguel Island, Azores
This alternative music and arts festival returns to the Azores’ largest island with a series of creative residencies. Its aim is to rethink Azorean heritage through the scope of music and art. Expect concerts, surprise moments, performative walks, talks, and more. This cool article from Conde Nast is a fun write-up if you’re curious.

May
North Music Festival
When: 24–26 May 2024 (2025 dates TBC)
Where: Parque de Serralves, Porto
In 2023 Robbie Williams anchored this music festival that promises a chilled vibe, great stages, a food court and entertainment throughout the day. In 2024 the first headliner announced is Keane, and it will be held at the beautiful Parque de Serralves.

June
Coala Festival
When: 31 May – 1 June 2025
Where: Hipódromo Manuel Possolo, Cascais – 30-minute train from Lisbon
This new-to-Portugal festival first made a splash in São Paulo, Brazil, back in 2014. The idea [cut off]

Requires one-hop browsing

SONAR LISBON 2025
LISBON, PORTUGAL
APRIL 11–13, 2025

MOGA FESTIVAL CAPARICA 2025
COSTA DA CAPARICA, PORTUGAL
MAY 28–JUNE 1, 2025

KALORAMA FESTIVAL 2025
LISBON, PORTUGAL
JUNE [illegible], 2025

SUNCEBEAT NEW HORIZONS 2025
LISBON, PORTUGAL
JUNE [illegible], 2025

NOS ALIVE 2025
LISBON, PORTUGAL
JULY 10–12, 2025 / BUY TICKETS

Interface text:
Confidential - Google DeepMind
Top 3 hotels in Lisbon
Dom Pedro Lisboa — Rating 8.6
Tivoli Oriente Lisboa Hotel — Rating 8.5
Hotel Roma — Rating 8.5
Explore all hotels
MORE INFORMATION
max
Premium
$20.99
LISBON
ALL GENRES
ALL VENUES
ALL MONTHS
ALL SIZES
NUMBER OF DAYS
ALL ORGANIZERS
MAP VIEW
CAMPING FESTS ONLY
VOLUNTEER
SORT BY POPULARITY
SHOW LINEUP POSTERS
RESET FILTERS
Your search has returned 6 festivals!
Cheer on your NBA team by grabbing verified tickets today.
See Tickets
ticketmaster
ADVERTISEMENT

[10:28](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=628s) [Two festival webpages contrast information available on one page with a listing that requires further browsing.](https://assets.ai.engineer/sites/aiecode2025/sha256/f9d17bacd8053c82e94f44c0fef5b8cfbe9989bf4fc4463404ffdb2b34aa7f59.webp)

Challenges: Web is Fragmented

Easy page to retrieve info

The best Portuguese music festivals in 2025
April 2025
Tremor Festival
[illegible]

May
North Music Festival
When: 24–26 May 2024 (2025 dates TBC)
Where: Parque de Serralves, Porto
[illegible]

Requires one-hop browsing

SONAR LISBON 2025
MOGA FESTIVAL CAPARICA 2025
KALORAMA FESTIVAL 2025
SUNCEBEAT NEW HORIZONS 2025
NOS ALIVE 2025
[illegible]

Interface text:
Confidential - Google DeepMind
Top 3 hotels in Lisbon
MORE INFORMATION
max
$20.99
LISBON
ALL GENRES
ALL VENUES
ALL MONTHS
ALL SIZES
NUMBER OF DAYS
ALL ORGANIZERS
MAP VIEW
RESET FILTERS
Cheer on your NBA team by grabbing verified tickets today.
See Tickets
ticketmaster
AIE
AGENT ENGINEERING
HTTPS://AI.ENGINEER

### Challenges: Context Management

Research outputs and follow-up questions can quickly fill the context window. The diagram moves older research notes into RAG, where they can be retrieved for a new task.

[10:41](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=641s) [A context window retains recent research tasks while notes from older tasks are stored for retrieval.](https://assets.ai.engineer/sites/aiecode2025/sha256/6436028ca5818fd660f7bd1b9d176e32bf5e54194472731de736af765ba48330.webp)

Challenges: Context Management

Context window

Task 1
Research query
Final report
Follow-up Q&A

Task 2
Research query
Final report
Follow-up Q&A

...
Research query
Final report
Follow-up Q&A

Task N
Research query
Research notes
Final report
Follow-up Q&A

N+1 task
Research query
Research notes

RAG
Research notes
Research notes
Research notes

Confidential - Google DeepMind

Interface text:
AIE
AGENT ENGINEERING
HTTPS://AI.ENGINEER

### Challenges: Context Management

Intermediate outputs and follow-ups grow the context. This design keeps more detail for recent tasks and stores older research notes in RAG, so the model can retrieve them selectively.

[10:51](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=651s) [A context window retains recent research tasks while notes from older tasks are stored for retrieval.](https://assets.ai.engineer/sites/aiecode2025/sha256/a9a6f6cfd7417a665cf27e56ae1e30602a0ef14669792c8bb20be5cf4c301a53.webp)

Challenges: Context Management

Context window

Task 1
Research query
Final report
Follow-up Q&A

Task 2
Research query
Final report
Follow-up Q&A

...
Research query
Final report
Follow-up Q&A

Task N
Research query
Research notes
Final report
Follow-up Q&A

N+1 task
Research query
Research notes

RAG
Research notes
Research notes
Research notes

Confidential - Google DeepMind

Interface text:
AIE
AGENT ENGINEERING
HTTPS://AI.ENGINEER

[11:37](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=697s) [Diagram showing recent research tasks in the context window and older research notes retrieved through RAG.](https://assets.ai.engineer/sites/aiecode2025/sha256/d185cf02ca4068a918740d40c13c486436d4db052f4dec70533af31458ea7fca.webp)

Challenges: Context Management

Context window

Task 1
Research query
Final report
Follow-up Q&A

Task 2
Research query
Final report
Follow-up Q&A

...
Research query
Final report
Follow-up Q&A

Task N
Research query
Research notes
Final report
Follow-up Q&A

N+1 task
Research query
Research notes

RAG
Research notes
Research notes
Research notes

Confidential - Google DeepMind

### What’s Next

The team compares today’s text-only, open-web research to a future with greater expertise, personalized research, and work that combines web research with coding, data science, and video generation.

[12:03](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=723s) [Comparison of current web research with proposed advances in expertise, personalization, and combining research with other capabilities.](https://assets.ai.engineer/sites/aiecode2025/sha256/9e19dfd79a5bdd8b27d116a232ef090d9ea1eea74280a8cc4c50f3f45530ea40.webp)

What’s Next

Where we are today:

McKinsey Analyst
• Retrieval from open web
• Text in/out only

Where we are heading

Expertise:
• Postdoc scientist
• “Partner” level insights(law, finance, consulting)

Personalization:
• Adapt the research to meet the user where they are

Multi-domain:
• Combine web research with coding, data science, video gen

Confidential - Google DeepMind

### What’s Next

Research agents could move beyond open-web, text-only research toward partner-level expertise, personalized research, and work combining web research with coding, data science, and video generation.

[12:09](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=729s) [Slide comparing current research capabilities with directions for expertise, personalization, and multi-domain work.](https://assets.ai.engineer/sites/aiecode2025/sha256/2ed1502d9369c14e476a9b73cd63573c9c1b27987d812b2df4ac196e0b3a6393.webp)

What’s Next

Where we are today:

McKinsey Analyst
• Retrieval from open web
• Text in/out only

Where we are heading

Expertise:
• Postdoc scientist
• “Partner” level insights(law, finance, consulting)

Personalization:
• Adapt the research to meet the user where they are

Multi-domain:
• Combine web research with coding, data science, video gen

Interface text:
Confidential - Google DeepMind
AIE
AGENT ENGINEERING
HTTPS://AI.ENGINEER

## Full transcript

[0:00](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=0s) [on-hold music] Hey, everyone.

[0:17](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=17s) I'm Aarush. I'm a product manager here at Google.

[0:19](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=19s) Hey, I'm Mukund. I'm a software engineer at Google working on Deep Research.

[0:22](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=22s) Um, so, uh, I don't know if people have had a chance to, uh, try Deep Research on Gemini, um, or are familiar with the product, but you can try it if you go to Gemini Advanced.

[0:34](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=34s) And if you scroll past 2.0 Flash, 2.0 Flash Thinking Experimental, 2.0 Flash Thinking Experimental with apps, 2.0 Pro Experimental, you will find, uh, 1.5 Pro with Deep Research, which is what we built.

[0:47](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=47s) Um, and if you have the chance to use it and you paid the twenty bucks, uh, you will see that it's a personal research agent that can browse the web for you to, to build, uh, reports on your behalf.

[0:58](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=58s) And so our motivation and what we wanna talk about today is kind of why we built it, some of the product challenges we overcame, and some of the technical challenges you'll face of building a web research agent.

[1:07](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=67s) Um, so our motivation was really we wanted to help people get smart fast. Um, we saw that research and learning queries are some of the top use cases in Gemini.

[1:17](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=77s) But when you bring, like, really hard questions, uh, to chatbots in general, what we were finding is that it would often give you a blueprint for an answer rather than actually give you the answer itself, right?

[1:29](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=89s) So we had this query that we used to throw around of like: Tell me what does it take to get an athletic scholarship for shot put, and, like, how do I go get one?

[1:38](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=98s) And often the answers would be things like, "You should talk to coaches," "You should find out how far you should be able to throw," and, you know, uh, "You should make sure you have good grades."

[1:46](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=106s) But really what I wanna know is like, okay, what are the grade boundaries? Like, how far do I need to actually be able to throw? I want something super comprehensive, and, and that's where we saw a big opportunity.

[1:56](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=116s) Yeah, so we said, 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.

[2:10](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=130s) But you gotta do it in five minutes, 'cause beyond that, uh, we don't have the chips. Um, so, uh,

[2:18](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=138s) this brought a bunch of product challenges for us. Um, Gemini, up to this point, is an inherently synchronous feature. It's a chatbot. Um, and so you wanted to-- we needed to figure out how do you sort of build asynchronous experiences in, in an inherently synchronous product.

[2:33](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=153s) Um, you also wanted to set expectations with users, right? Deep Research is good for, like, one very specific thing, but a lot of user queries to Gemini are things like, "What's the weather?"

[2:40](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=160s) "Write me a joke." Things like that, where waiting five minutes is not gonna get you a good answer, and we wanted to set expectations. Uh, and the last thing is our answers can be thousands of words long, and we needed to figure out how do you make it easy for users to engage with really long outputs and,

[2:56](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=176s) um, uh, in, in a chat experience. Um, so let's walk through kind of the UX and kind of think about how, how we solve some of these, right? So imagine you're a VC, uh, and everybody's talking about, you know, investing in nuclear in America.

[3:11](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=191s) And so you come with this query like, "Hey, help me learn the latest technology breakthroughs in small nuclear reactors and tell me interesting companies in the supply chain." So the first step, um, when you bring this query to Deep Research is that Gemini will actually put together a research plan for you and present it in a card.

[3:26](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=206s) And so this is the first way in which we're able to communicate with users. Like, this is different. This isn't your standard chatbot experience. Something's gonna happen. You're gonna hit Start.

[3:35](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=215s) But it's also an opportunity for us to actually show the user a research plan that they can edit and engage with. Kind of like a good analyst, right? They, they wouldn't just get to work.

[3:42](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=222s) They'd actually show you, "Okay, here's how I'm gonna approach this." And it's a way for users to, if they want, kind of engage and steer the direction of the research further.

[3:51](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=231s) Now, uh, once it-- you hit Start, we actually try and show you, um, what Gemini is doing on the-- under the hood in real time, uh, by showing you the, the websites it's browsing.

[4:03](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=243s) And this is a feature that was built before thinking models, and thoughts are also a really great way of kind of showing transparency of what the model is thinking.

[4:11](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=251s) Um, but what's really nice here is while you wait, you can sort of click through the websites, dive into any of the content. Um, but what we also inadvertently saw is people trying to game that number to see how high it could go.

[4:22](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=262s) So we definitely saw people push that number into the, into the thousands, uh, to try and, um, you know, see how many websites Deep Research could read.

[4:30](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=270s) Um, finally, we kind of get this report that's, you know, thousands of words long. And, um, we're really inspired by what, kind of what Anthropic does with, um, Artifacts.

[4:40](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=280s) And so we thought that was a really great way of sort of being able to pin an artifact so that users can actually ask questions about the research while reading the material.

[4:49](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=289s) They don't have to scroll back and forth. And what's really neat about this is it means it's easy for you to engage in sort of changing the style of the report, adding sections, removing sections, asking follow-up questions, and, uh, and it sort of makes that really easy.

[5:02](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=302s) And the last part that's super important is kind of user trust and also doing right by the publishers. So we, we try and always show is all the sources we read as well as all the sources we used in the report.

[5:12](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=312s) 'Cause not everything that we read is used, but it stays in context for follow-up questions. And, and also sort of these are all things that, um, carry over to Google Docs as citations and things like that if you choose to export.

[5:29](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=329s) Uh, so I thought today we can pick some of the challenges, uh, that one has to encounter while building a research agent and ta-talk through some of them. So, uh, I picked four for today.

[5:39](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=339s) So one is this, this long-running nature of tasks introduce-- This is a couple of things that we need to look into. Second is the model has to plan iteratively and spend, uh, its time and compute during this time effectively.

[5:55](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=355s) So what are those challenges there? And it has to do this, uh, while interacting with A very noisy environment that is the web. And as you do this and, uh, read through information, very quickly you can start seeing your context grow and h-how do you effectively manage context.

[6:14](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=374s) So if, if you think about a job that runs for multiple minutes and something that can make many, many d-- uh, different LLM calls and calls to different services, there are bound to be failures, right?

[6:26](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=386s) And today we are talking about, order of minutes, but you can very easily think in the future of, uh, these kind of research agents taking like multiple hours. So it's important to be robust to intermediate failures of these various services of various reliabilities.

[6:41](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=401s) And so being able to build a good state management solution, being able to recover from e-errors effectively so that you just don't drop the whole research, uh, task due to one failure.

[6:53](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=413s) That's one. The second aspect of doing this, what it enables us, is to enable this feature, uh, cross-platform. So we believe more and more, uh, users will start kind of registering your asks, uh, or your research tasks and just like walk away, do their thing, and then you need to get notified.

[7:11](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=431s) And this can happen now across, uh, devices and you can pick up, uh, uh, reading it, uh, uh, uh, once it's done.

[7:21](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=441s) So now what is the model doing at, uh, like through these, you know, uh, few minutes? Uh, so let's take, uh, example, right? So here, uh, we're looking for, uh, athletic scholarships, uh, for shot put.

[7:33](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=453s) There are many facets to this query, and we kind of show this in a research plan like Aarush showed. 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, right?

[7:49](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=469s) So the model has to be able to reason to do that. And, uh, the other challenge is, here you see you're always gonna land in this state where there's partial information, so it's important to look at all the information found so far before you decide what to do next.

[8:06](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=486s) So in this instance, the model found, hey, it's-- it knows the qualifying standards, uh, for the D1 division, but in order to provide a complete report and answer the user's question, it has to go figure out what the equivalent for the D2 and D3 divisions are.

[8:22](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=502s) So this notion of being able to ground on information you find and then plan your next step is key.

[8:31](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=511s) Another example of partial information could be when you make searches. Uh, so in this case, we're trying to find the best roller coaster, uh, for kids. Uh, you might find results, uh, that provide partial information again.

[8:44](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=524s) So here, uh, you end up at a link, uh, which talks about the top ten roller coasters, but does not mention anything about them being suitable to kids. Uh, so the plan has to recognize this fact and then go ahead and in the next steps of planning, try to resolve this, uh, disambiguity.

[9:05](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=545s) Um, another example of, uh, challenges in planning is information is often not found in one place. You find facets of information spread across different sources. So here, uh, we are trying to find, uh, what would, uh, what would it take to get a certification for a scuba dive, uh, in, in, in some dive centers nearby.

[9:26](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=566s) So you see, uh, one part or one source has, uh, the kind of the structure of, uh, what, what you have to go through to get a certification, but in a completely different source, you have this notion of the pricing for this diving center.

[9:40](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=580s) So the model has to weave this together to figure out, um, you know, what the cost structure for such a certification would look like.

[9:48](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=588s) Then there's the classic, uh, entity resolution problem. So you might find mentions of the same entity across different sources, so you need to be able to reason about some information indicators to kind of figure out if they're talking about the same entity or you need to explore more to verify such, uh, disambiguities.

[10:09](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=609s) Um, yeah, I think m-most people here have worked on some notion of a web problem, and we know like it's super fragmented. So, uh, here you see two different websites, uh, talking about the same thing, uh, about music festivals in Portugal this year.

[10:24](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=624s) Uh, on the left, uh, if you end up at such a website, it's easier and you get most of your information in one go. Uh, on the right, uh, the layout is different, so having a robust, uh, browsing mechanism if you wanna navigate, uh, the web for your research tasks is another, uh, important challenge.

[10:44](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=644s) So like we saw, there is a lot of these intermediate outputs, and as you do this and you start getting streams of information during your planning, you can imagine your context size growing very quickly.

[10:57](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=657s) Um, the other challenge that, uh, about context size is your research task doesn't typically end with your first query. People have follow-ups. People can say, "Hey, can you also do the same for this other topic?"

[11:10](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=670s) So there is like this kind of a follow-up, uh, deep research and, uh, that also adds pressure on the context. Uh, we at Gemini have, uh, the liberty of really long-context models, uh, but, uh, even then you have to design, uh, some way to make sure you, you effectively manage your context.

[11:30](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=690s) And there are multiple choices here, each come with various different trade-offs. Uh, we're showing one here, uh, where we kind of have like this recency bias. So you have a lot more information about your current and your previous tasks, but as you get to older tasks, we kind of selectively pick out, uh, you know, things what we

[11:50](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=710s) call as research notes and put it in a RAG. That way, the model can still access it, but it's being selective. Uh, I'll hand it back to Aarush about, uh, to talk about what's next.

[12:00](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=720s) Yeah. So we were super excited To put this feature out in December, we weren't actually sure if anyone was gonna use it, if anyone was gonna care, um, uh, to wait five minutes, uh, for something.

[12:12](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=732s) And, uh, we were really positively surprised by the reception. Um, and, and really what we, what we saw, um, was, hey, we, we've built something that's maybe as good as, like, a McKinsey analyst, right?

[12:23](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=743s) And we give it away for 20 bucks. But, um, you know, that's, that's really great and, uh... But what it does is it just retrieves from the open web, and it's a text in, text out only system, right?

[12:34](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=754s) And so where we sort of, we sort of see a few different directions of where research agents are gonna go next, and the first one is around expertise, right?

[12:42](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=762s) So how do you go from a McKinsey analyst to a McKinsey partner or a Goldman Sachs partner or, like, a partner at a law firm, right? So that's really around not just being able to aggregate information and synthesize it, but also think through the "so what" of how do-- like, what are the implications for what we're gonna

[12:58](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=778s) do and, and what are the most interesting insights and patterns that come out of it? The, the other thing is, you know, there are plenty of domains beyond professional services, like the sciences, where you, you know, wanna get really good.

[13:09](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=789s) You know, you want something that can read many papers, form hypotheses, find really interesting patterns in, you know, what methods were used, uh, and, and come up with novel hypotheses to explore.

[13:20](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=800s) However, um, just because you build something that can be really smart doesn't mean that it's useful to someone, right? So, um, if we were thinking about a use case of running a due diligence on a company, the way you'd present that information to me would be very different to the way you'd present that information to, say, a

[13:36](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=816s) Goldman Sachs banker, right? Um, for me, you really wanna talk through, like, what, like, what is this company and how is it positioned strategically? But a banker would want to know all the financial information, actually have a DCF that they could look at, right?

[13:50](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=830s) Actually, uh, have a, have a much more, like, fine-grained, uh, sort of, uh, finan- uh, financial modeling and analysis. And, and that really should shape the way in which you browse the web, right?

[14:00](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=840s) 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.

[14:07](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=847s) I think the last part is sort of something that goes across domains of what models can do, right? So not just being able to do web research with text, but being able to combine that with abilities in coding, data science, even video generation, right?

[14:19](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=859s) So coming back to this example, if you're doing a due diligence, y- what if it could go and do, like, a lot of statistical analysis and actually build financial models to inform the research output that it gives you, right?

[14:29](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=869s) Telling you, "Hey, why is this a good company or not?" Um, I should say Google doesn't give financial advice and- [laughs] ... you know, it's not a financial advisor. Um, but yeah.

[14:39](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=879s) And so we're really excited about the potential. We think there's a ton of headroom to make research agents better, and we are really glad we didn't call this Gemini Deep Dive, which was [laughs] our best name before, uh, before launching this feature.

[14:51](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=891s) Um, that's it. Thank you so much.

[14:54](https://www.youtube.com/watch?v=eJOjdjO45Sc&t=894s) Thank you. [clapping] [upbeat music]
