AI Agents as Research Assistants: Records, Links, and History
Published:
Stop explaining your research context from scratch
When you ask generative AI about your research, do you repeat what you have checked, which sources support your work, and what direction you are taking? When notes and materials are scattered, both you and generative AI have trouble tracing earlier decisions.
An AI agent can read research-note files and instructions, then work across several records. But giving it scattered information does not make it understand your decisions correctly. A research foundation of records, links, and history helps both you and the agent check evidence and the current direction.
What you will learn
- Which tasks to give an AI agent, and which decisions to keep for yourself
- Why records, links, and history form a research foundation
- What to confirm before giving research materials to an AI agent
- How to develop an existing set of notes into a research foundation
AI agents can work across records
Here, an AI agent is a way of using generative AI that can work with files, tools, and continuing instructions across several steps. It can read research notes, find related records, suggest what to check next, and draft proposed changes.
This differs from chat-based generative AI used for a single question or summary. An AI agent should not decide your research question, conclusion, choice of method, or what information may be made public. Use it as a research assistant: to find evidence, read records, and suggest missing points or alternatives. You decide what to adopt and what to claim.
The Research Foundation and AI Use Guide will explain the difference between chat-based generative AI and AI agents, and how to choose where to start.
A research foundation consists of records, links, and history
A research foundation is not the name of a particular app or service. It is a way to let both people and AI agents trace evidence, decisions, records, and changes.
Start with these four elements.
| Element | Role | Why it matters when using an AI agent |
|---|---|---|
| Records | Keep verified facts, working notes, open questions, and decisions in separate roles | Makes it less likely that unfinished thoughts will be treated as evidence |
| Markdown | A readable text format with structure from headings and lists | Makes it easier to read the needed parts and propose explanations or changes |
| Links | Connect evidence, related records, and decisions | Shows which sources and records to check |
| History | Keep track of changes and compare before and after | Lets you review an AI agent’s changes and revert them if needed |
In Part 6, I introduced verified facts, working notes, open questions, and decisions. Keep sources with verified facts, interpretations and hypotheses in working notes, unresolved questions in open questions, and the current direction in decisions. This separation makes it less likely that you or an AI agent will confuse a current decision with an unfinished thought.
Markdown is a text format that you can write without a special note-taking app. Generative AI often uses Markdown in prompts and responses. Keeping your notes in the same format makes headings, lists, and links easier to handle. It also helps people read the notes again and helps AI agents find the needed parts. The Research Foundation and AI Use Guide will cover basic Markdown.
Links are not decoration for collecting many materials. For example, link a decision to the working note that explains it, then link that note to the paper or experiment record that supports it. This lets you check not only a conclusion, but also the information behind it.
History also makes changes less risky. When an AI agent proposes edits, you can compare the before and after versions before accepting them. If needed, you can return to an earlier version and try a small improvement again.
Choose tools by their role
Obsidian is a tool for viewing and editing Markdown files. Its linking features make related records easier to follow. Git lets you compare file changes, keep a history, and return to an earlier state. I use this combination.
However, Obsidian and Git are not the only way to conduct research. If your current note-taking and storage method can preserve the role of each record, a path to its evidence, and a way to check changes, it can be your starting point. The important point is not the tool name. It is whether you can trace evidence and decisions later.
The Research Foundation and AI Use Guide will cover how to start with Obsidian, Markdown, Git, and GitHub. This article focuses on what the tools need to support before you add them.
What people need to confirm before asking an AI agent to work
Before giving research materials to an AI agent, confirm at least these four points.
- Reading scope and restrictions: Which folders and documents must the agent read first? Which must it not read? Explicitly exclude unpublished materials, personal information, or materials you are not allowed to give to an external generative AI service. In all other areas, let the agent follow related records and links.
- Roles of information: Do not mix verified facts, working notes, open questions, and decisions. An AI agent may propose categories, but you check the sources and content and decide how each record will be used.
- Scope of changes: Will you ask only for reading, explanation, or suggestions, or also for editing and command execution?
- How to check: How will you check sources in the output, changes before and after, and whether anything will be sent outside your workspace?
At first, specify the materials the agent must read and the areas it must not read, then ask only for reading and suggestions. Within those boundaries, following related records and links can reveal relationships or missing information that the initial documents alone would not show. Before allowing edits or command execution, check the purpose, target, and expected changes.
The Research Foundation and AI Use Guide will cover practical ways to separate materials, limit permissions, and check changes.
Develop existing records into a research foundation
If you are already in the middle of research, you do not need to rebuild every record from the beginning. Choose one current note or piece of material. Separate verified facts, working notes, open questions, and current decisions. Add links to its sources and related records. An AI agent can suggest categories and links, while you check the content.
The Research Foundation and AI Use Guide will show the practical steps for using an AI agent to divide existing records, add links, and move them into a research foundation while checking changes. A research foundation does not need to be complete at once. Build the parts you need as your research continues.
Summary: Build a research foundation with AI agents
You do not need to finish organizing every record before using an AI agent as a research assistant. Let the agent suggest categories, links between records, and possible rules while you develop the foundation step by step.
You remain responsible for identifying what the agent must read and must not read, distinguishing facts from interpretations, and deciding whether changes should be accepted. Markdown, links, and history help you develop the research foundation with an AI agent while keeping that review possible. Part 8 will cover how to avoid settling a decision too soon based only on a plausible explanation from generative AI or an AI agent.