Meet the AI Research Agent: research you can check, step by step

Ask a general AI chat tool for sources on your research topic and you may get a confident list of papers, complete with authors, years and journal names. Some of them may not exist. Others may exist but not say what the tool claims they say. For a student writing a long essay, a project or a thesis, that is worse than no help at all: a made-up reference can cost you marks, or raise questions about academic honesty.

The AI Research Agent in Studentlytica AI is built around one rule: every statement it makes about the literature must point to a real paper and to the exact words in that paper that support it. This article explains what the agent does, how it works, and why that rule matters. It uses one example project throughout.

What the AI Research Agent does

The AI Research Agent helps you find, read, compare and write from academic papers, inside a private research project. A project holds your research question, the papers you have found or uploaded, your notes, the agent's reports and your own writing. Its main jobs are:

• Finding papers in an open scholarly catalogue.
• Running deep research: a planned, multi-step search and reading of the literature on your question.
• Reading and comparing papers: what each paper studied, how and what it found, and where papers agree or disagree.
• Answering your questions about the papers in your project.
• Writing a literature review of the papers you select.
• Helping you write, one section at a time, with citations and a reference list.
• Checking your references and claims: are the papers' details right, has a paper been retracted, and does the evidence really support each statement?

Why "agent"?

An AI agent works towards a goal over several steps and decides what to do at each step. For deep research, that means the agent plans the search, runs it, decides which papers are worth reading, reads them, notices what is still missing, searches again for that, and finally compares what it found. A single chat answer cannot do this.

But a research agent that decides everything by itself would be hard to trust. So the important decisions are made by code you can inspect in the report:

• Retracted papers, papers outside your chosen years and papers with no text to read are excluded by code, not by the model's judgement.
• The number of searches and papers is fixed by your plan, and the run stops and tells you if it reaches its limit, rather than carrying on quietly.
• Every quotation the model gives must be found word for word in the paper's text, or the statement is rejected.

An example project

The example below illustrates how a project flows. The student and the project are made up for this article, and no real papers or findings are quoted. In the app, every paper and quotation comes from the real literature and your own uploads.

Kwame is a final-year economics student. His project question is: How has mobile money changed household savings in Ghana? He creates a project, adds the question and the years he is interested in, and starts a deep research run.

1. The plan

The agent first turns the question into a short plan: a few sub-questions (for example, who saves through mobile money, and does mobile money replace or add to other savings), the key concepts, what a relevant paper should cover, and a handful of short catalogue searches. Only those short search terms leave Studentlytica AI. Kwame's project notes and files do not.

2. Searching and screening

The agent runs the searches in the catalogue and screens what comes back. Code removes retracted papers, papers outside Kwame's years and duplicates found by several searches. The model then sorts the rest by title and abstract into relevant, maybe and not relevant, and code picks the strongest candidates. A paper the model judged not relevant is never chosen.

If one of the sub-questions is not covered by any chosen paper, the agent searches once more for that sub-question alone. If it is still not covered, the report says so.

3. Checking the papers are real and current

Before adding papers to the project, the agent checks each paper that has a DOI (the permanent identifier most journal articles have) against Crossref, the registry publishers use. A paper the publisher has retracted or withdrawn is left out, and the report explains why. Corrections and expressions of concern are shown, not hidden.

4. Reading

The agent then reads each chosen paper on its own. Where a paper is openly licensed and its publisher allows downloads, Kwame can let the agent save the full text into one of his courses, and the agent reads the whole paper. Otherwise, it reads the abstract, and every statement based only on an abstract is labelled that way. The agent never bypasses a paywall or claims to have read a paper it could not open.

For each paper, it records what the paper studied, its methods, its findings and its limitations, each as a statement linked to the exact sentences it came from.

5. Comparing

Finally, the agent compares the papers. Each statement across papers gets a strength label worked out by code from the evidence behind it:

• Strong: three or more papers, at least two of them read in full.
• Moderate: two or more papers.
• Limited: one paper.
• Contested: at least one excerpt in the project says otherwise. Those excerpts are shown under "says otherwise", so Kwame can read both sides.

Strength describes how widely a statement is supported in the papers found, not how good the studies are. Judging study quality is still Kwame's job, and his supervisor's.

6. The report

The finished report lists the plan, every search, every paper found and what happened to it (chosen, left out and why, read in full or from the abstract), the statements with their strength, and the sub-questions the papers did not answer. It says clearly that it is not a systematic review: it is a bounded search of one catalogue, not a complete survey of everything published.

Next to any statement, Show evidence opens the exact passages behind it.

Working with your papers

The report is a starting point. Inside the project, Kwame can:

• Search for more papers himself, and add or upload papers he already has. Each paper shows what is known about it: its catalogue details only, its abstract or its full text.
• Ask questions about his papers, or about one paper in the Paper Reader. Answers quote the papers and say when the papers answer only part of the question. Other question modes help him explore the topic and refine it into answerable research questions, list provisional gaps in these papers, or compare the methods the papers used and suggest methods he could use, to discuss with his supervisor and his institution's ethics requirements. Suggestions are labelled "Suggested by Studentlytica, not taken from your papers".
• Build an evidence table across the papers.
• Ask for a literature review of the papers he selects: up to six themes, each point cited, each theme with its strength, and a clear statement of what was searched and what was not.

Writing, with your voice and your responsibility

In the writing workspace, Kwame creates a document with his own sections, such as Introduction, Literature review and Methodology. He writes in it himself and cites papers from his project. The citations and the reference list are formatted in his project's citation style, and he can export the document to Word, Markdown or plain text.

He can also ask the agent to draft one section from the evidence in his project. The draft is written only from excerpts of his papers. Where the section needs something the papers do not show, the draft does not invent it: it adds a visible marker such as [Needs evidence: recent figures on rural mobile money use], so Kwame knows what to find. The draft is kept apart until he chooses Insert, and the document then notes which sections include drafted text.

The agent drafts sections, never a whole thesis in one request. You remain the author of your work and responsible for following your institution's rules on AI use and disclosure. Many institutions now ask students to say how they used AI tools. Check yours.

Checking your references and claims

Two checks help Kwame trust what he submits:

• Reference check. For papers with a DOI, the agent compares the title, year and first author with Crossref's record and reports each paper as verified, mismatch (showing what differs) or not found. It also shows any retraction, withdrawal, correction or expression of concern reported for the paper. "Not found" is not proof that a paper is fake, and nothing is changed without Kwame's say.
• Claim check. On request, each statement is checked against the passages it cites: does the passage support it, partly support it, or not support it? Code also flags common problems, such as a number in the statement that the passages do not contain, cause-and-effect wording when the papers only report an association, and a statement about "the literature" that rests on a single paper. The check can be wrong, and the app says so. It is there to point Kwame to the statements he should read again.

Why this matters

Research is a skill, and a large part of it is knowing exactly what each source does and does not say. The AI Research Agent saves time on the slow parts, such as searching, screening, extracting and organising, without hiding the evidence from you:

• No invented sources. Papers come from a scholarly catalogue or from your own uploads, and references are formatted from each paper's recorded details, never written by the model.
• Every statement is traceable to the exact words in a paper.
• Honest about what was read. Abstract-only reading and papers that could not be read are labelled, not glossed over.
• Honest about coverage. Searches are disclosed and limited, and the report never claims to be a systematic review.
• You stay in charge of the question, the papers, the argument and the final text.

No tool can make a piece of research complete or correct for you. What the AI Research Agent can do is help you build your work on evidence you can see and check.

Practical details

• Allowance. Deep research runs, questions, literature reviews and section drafts each use your monthly allowance; claim checks do not. See Pricing for each plan. Pro and Max runs search and read more widely than Free runs.
• Privacy. Your projects, notes, uploads and writing are private to your account. Catalogue searches send only short search terms; reference checks send only the DOI. To read and write, Studentlytica AI sends our AI provider only the paper passages needed for that step and your question, never your name, email address or account details. The Privacy Notice has the details.

Try it

Open the AI Research Agent, create a project with your research question, and start a deep research run. Then read the report the way a careful researcher would: open the evidence behind the statements that matter most, look at what was read only from abstracts, and note what the papers did not answer. That is where your own research begins.