Getting Started with AI Agents: A Guide to using AI for Research

In 2025, AI tools are advancing beyond simple question-and-answer chatbots. Now, multi-stage tasks can be delegated to AI models that will produce and execute plans to create entire research reports. This marks the rise of AI agents. But what is an AI agent, and how could you use them?

‍

In the most popular AI textbook "Artificial Intelligence: A Modern Approach," Stuart Russell and Peter Norvig define an agent as an entity that perceives and manipulates the environment it operates within. Put simply: agents are entities (not necessarily AIs) that make decisions and take actions.

Tobias Zwingman elaborates on the spectrum of autonomy that these agents may have. The simplest are 'reactive agents' - think of a Roomba, which only acts when it bumps into something or when its battery is low. Meanwhile, the most 'agentic' AI systems that major labs are aiming towards can be classified as 'learning agents,' which take actions and iteratively improve by observing the results.

With the definitions out of the way, let's consider how agents could be introduced into your firm's operations in 2025.

‍

The Case Study of the Research Assistant Agent

‍

We have entered an era of semi-independent research agents.

Consider tasking an employee to produce a briefing on a new area for your firm. To produce a quality report, they need both sufficient writing skills and the ability to scour through research papers while maintaining focus. This dedicated, multi-step process can now be completed by Gemini Advance 1.5 Pro's deep research mode (research mode henceforth). A well-cited, tailored 10 page report, which used to take days of employee time, has dropped to near-zero cost. 

Suppose you're an executive who needs to familiarize yourself with an emerging neighboring market. Imagine a logistics executive who has received a pitch from a drone delivery startup. The opportunity may be lucrative, but exploring it will take valuable time. Research mode will near-instantly provide you with the data to qualify-in or -out this opportunity and whether it's worth exploring further. 

In this case, less than 3 minutes of human labor and a 5-minute wait resulted in a tailored report. The 10-page document is the result of analyzing 100+ webpages and it includes 30 citations ranging from regulators' internal pages to PwC forecasts. Research mode also introduced additional considerations such as the role of crime and theft in the market. It even helpfully mapped out the drone-startup space to inform you of the pitching firm's competitors.

What's particularly impressive about Research Mode is its capacity to produce multi-step plans that anticipate key considerations. Our original prompt did not mention the regulatory environment, yet the produced draft identified the key licenses and agencies required for commercial drone operation in the UK. You can see our conversation on evaluating the UK drone market here. 

‍

The Gemini Advance bundle that includes the Research Agent has a free month trial. This gives you time to identify use cases and hone your prompting skills before committing to a subscription.

‍

How to Use Gemini's Research Mode

‍

The user interface of Gemini's chatbot will be familiar to you from your previous conversations with LLMs and yet the range of outcomes in Research Mode are enhanced through the multi-prompt 'planning' phase when delegating a research task. The first prompt of the conversation leads to a joint brainstorming session between you and the chatbot to produce a comprehensive plan. You can direct which webpages and forums should be emphasized, set the output format, and even prioritize the data types that are worth scraping. No more searching for emails and phone numbers in the 'contact us' subpage required.

Here's how to use research mode:

1. First, open your Gemini page and select Gemini Advance 1.5 Pro with Deep Research from the top-left menu, or the future equivalent model labeled with the Deep Research capability.
2. Then, describe your query. For example: "I'm evaluating the viability of implementing drone delivery services for consumer products in London and South East England through a potential startup partnership. Please provide a detailed analysis covering…” 
3. Next, Research Mode will define a multi-step plan, articulating which questions it will investigate, what sources it may use, and the general structure of its output report.
4. At this stage, you can specify and modify your requests, asking it to target particular sources, emphasize scraping contact information into a table, or add whatever other features you need from the report.

‍

The researcher can achieve impressive accuracy with minimal context. All of its claims are cited directly from sources and can be independently verified. Within a few queries, you'll develop a clear understanding of which questions Research Mode handles best.

This referencing capacity addresses a frequently asked question by our partners and clients: what about hallucinations? All content claims are made entirely on existing sources. Every claim has its source footnoted with the access link placed in the bibliography of the document. Hallucinations are no longer a major issue, yet all diligent research involves verifying claims. 

Research Mode's existing limitations are (of course!) tied to the general constraints of online desk research, namely the quality of source material. Search queries that have been overloaded with SEO techniques may risk repeating advertising copy, such as in the case of nutrition topics. However, this can be mitigated by explicitly requesting reports based partially or entirely on academic papers or even on the collective wisdom of specific forums.

We can only anticipate that the accuracy, complexity, and depth of AI agent capabilities will increase over time. While the first tasks to be semi or fully automated will be procedural ones, Research Mode has demonstrated that even complex judgments and analyses are within reach of current AI technology. As one newsletter we like described it: we have achieved “the industrialisation of basic research!” 

‍

AI labs are focused on increasing the level of autonomy, planning capacity, and range of tasks that AI agents will be capable of undertaking. As a result, a new skill of AI agent management is emerging. Those who can leverage the rapid deep work of AI will be able to massively increase their firms' output, cutting day-long tasks down to minutes. Thanks to the 1 million token context window and the upcoming ability to upload documents to Gemini Research mode, identifying patterns from internal data will become a matter of typing out a few sentences. Familiarize yourself with the agent and get ahead of the curve.

‍