Using Generative AI for Research: Chat-Based AI and AI Agents
Published:
Conclusion: I use AI agents by default, but you can start with free chat-based AI
I use AI agents as the default in my research because I want discussions to connect to context, evidence, and records. It is easier to work when the agent can reach papers in my files, my current hypotheses, past decisions, and rules about material it must not read.
Many of the practices in this article also work with free chat-based generative AI. You can type or attach the relevant facts and your current thinking, check the evidence, and return the outcome to your records. AI agents and paid plans are not requirements for starting research.
This article is for readers who have used chat-based AI such as ChatGPT or Claude and want to understand the role of AI agents in research. You do not need command-line or Git knowledge to follow it.
“Use chat for simple tasks” is not enough
People often distinguish chat-based AI and AI agents by whether a task is light or heavy. That distinction is not enough in research, because even a task that seems simple may need to connect to background and evidence.
For example, consider asking for an explanation of an unfamiliar term or idea. A general explanation may be enough in a chat. But if you want the explanation to cite the paper you are reading or material in your research notes, an AI agent that can reach those sources is easier to use.
To ask for the same thing in a chat, you need to select and provide the papers or records you want cited each time. You also need to decide which material, and how much of it, is enough for the explanation. If you leave out a needed source, the evidence and scope of the comparison become narrower.
The same applies when you organize what to check before reading a paper. In a chat, you provide the research question, your current hypothesis, and what you have already confirmed as needed. With an AI agent, you can tell it to refer to organized records and links, reducing the need to explain the same background every time.
The point is not that an AI agent always gives the right answer. It is that you can trace which sources support its explanation, separate facts from suggestions, and make your own judgment more easily.
AI agents make it easier to carry context forward
An AI agent is a way of using generative AI that can work with files, tools, and continuing instructions across multiple tasks. In research, it helps to keep materials and records separated by purpose, with links between the things that need to be connected.
In my own workflow, I specify files containing verified facts and files containing my hypotheses or current thinking, and I link related records. I also set rules for material that must not be read or sent outside. This means I do not need to explain my research background from scratch every time I want to discuss something.
Part 7 introduced how to bring AI agents into research. Providing needed context in a chat and making context reachable through records for an AI agent are different methods, but the same principle applies to both: do not mix verified facts with hypotheses or ideas.
ChatGPT Work (an AI agent for the general-purpose work I previously did with Codex) can be used for longer research, information analysis, and creating deliverables such as documents. Coding AI agents such as Codex and Claude Code are suited to technical work with local files, repositories, and terminals. Features and environments change, so check each service’s official guidance before starting. See the official guide to ChatGPT Work and Codex and the Claude Code documentation.
I use Chat to allocate usage, not because it is less capable
I still use chat-based generative AI. I do not use Chat because I think certain tasks are too small or unimportant for an AI agent. In my ChatGPT account, Chat and Work have separate usage allowances, so I want to save my Work usage for tasks such as reading research materials and doing longer research or organization.
I therefore use Chat for tasks that need no additional research context, or for tasks where I can provide the needed information briefly. This is not because Chat is inherently a better fit. It is a way to decide where to spend limited usage.
This allocation will not apply to every reader. It depends on the service, subscription, usage limits, and how you keep research materials. Rather than treating Chat and AI agents as separate rankings, use the context you need and the usage you have available as the basis for your own workflow.
You can practice the same research habits with free chat-based AI
You do not need to change how you approach research if you do not use an AI agent or are not ready to subscribe. In the chat input, provide the current question, verified facts, your hypothesis or idea, and what you want to check. If you can attach material, state which part to read and what you want to verify.
Ask not only for confirmation, but also for counterexamples, alternatives, and evidence that still needs checking. After the conversation, separate the content into verified facts, working notes, a provisional decision, and the next question. These practices are the same with chat-based AI and AI agents. Part 8 and Part 9 explain this checking and recording process in more detail.
AI agents are an option for reducing the effort of providing context and returning work to your records. Start with free chat-based AI, then consider an AI agent when you find yourself repeatedly explaining the same background, wanting to check related material together, or needing a proposed change across multiple files.
What remains the human’s responsibility
Chat-based AI and AI agents differ in the context and range of work they can handle. The human responsibilities in research do not change.
- Decide the research question: what you need to find out.
- Decide which material the AI may read and what must not be sent outside.
- Check the output against original sources and records.
- Review file changes and execution results before accepting them.
- Make decisions about conclusions, research direction, and scope changes.
Generative AI can be a research assistant that surfaces possible evidence, counterexamples, and blind spots. But adopting a plausible option without checking it can end the investigation too early. Check the necessary properties, assumptions, alternatives, and evidence before deciding.
Summary
In my research, I use AI agents by default so that discussions can connect to context, evidence, and records. I use Chat not because it is less capable, but to allocate limited usage.
Still, free chat-based AI can support the same research practices when you provide the needed context, check evidence, and return the outcome to your records. Start with the form you can use now, and consider an AI agent when the need to carry context forward becomes important.
A related guide will cover how to decide whether to subscribe to ChatGPT, Claude, or Gemini.