What a research agent actually is.

From “summary” to “research” CORE

You have asked an AI to summarize something before. You typed “explain quantum entanglement” or “give me the arguments for and against a four-day work week” and the model produced a few paragraphs. That output looked like research. It was not.

What you received was the model’s compressed memory of its training data, phrased as a summary. The model did not go read any sources. It did not check whether its training was accurate. It did not look at anything newer than its cutoff date. It gave you an average of what it had seen, stated in a confident voice. That is a summary, and for many tasks — explaining a term, giving an overview — a summary is a perfectly fine output.

A research agent is a different kind of thing. It goes and retrieves. It runs searches against the live web, opens pages, reads documents you give it, follows links, and brings actual sources back into the conversation. The output is anchored in specific, nameable places — not an average of everything it has ever seen.

The difference matters for three reasons. First, freshness: a research agent can tell you something that happened last week; a summary cannot. Second, verifiability: a research agent’s claims are traceable to sources you can open; a summary’s claims are traceable to nothing. Third, correctability: when a research agent is wrong, you can find out by checking its sources; when a summary is wrong, you typically cannot tell.

This is the same category shift you learned in Module 3 — from suggester to coding agent. There, the jump was tool use: the agent could read files, run commands, verify results. Here, the tool is web retrieval (and sometimes file-reading tools pointed at your local sources). The principle is identical. An agent that can reach outside itself is a different category from one that cannot.

The four moves of a research agent CORE

Every real research task a directed agent performs runs through four moves. Naming them gives you the handles you need to direct the work and check it.

Scope. The agent — or, much better, you — figures out what question is actually being asked. Not “tell me about X” but “I want to know Y about X, for purpose Z, and here is what would count as an answer.” A huge fraction of bad research-agent output is the result of a scope that was left vague. Scoping is almost entirely a human move; the agent can help, but you are the one who knows why you are asking.

Retrieve. The agent fetches candidate sources. It runs web searches, fetches pages, opens documents. A good agent does this across a range of source types — academic, journalistic, primary-document, domain-specialist — rather than grabbing whatever is at the top of a single search. A bad agent grabs the first four Google results and calls it done.

Triangulate. The agent compares what those sources say. Which claims appear in multiple independent sources? Which appear in only one? Which sources disagree, and why? Which are primary and which are secondary? Triangulation is where research stops being a pile of links and starts being an answer that can be defended.

Synthesize. The agent turns the triangulated findings into a structured output — a brief, a memo, a comparison, a paragraph of prose — that a human reader can use. Synthesis is where clarity happens, and also where the biggest honesty test lives: does the output represent what the sources actually say, including their uncertainty, or does it smooth everything into confident-sounding prose?

As the director, you check three of those moves and read the fourth critically.

You own scope. The agent does not know why you are asking; you do. If scope is wrong, everything downstream is wrong, no matter how well the agent runs.

You audit retrieve by asking: what sources did it actually fetch? Are they the right kind? Is anything important missing? Is any of it fabricated? You look at the list, not just the summary.

You check triangulate by reading enough of the sources yourself to confirm the agent’s comparisons. This is the single move most likely to go wrong invisibly — the agent can claim two sources “agree” when they are addressing slightly different questions, or claim a lone source is “the consensus” when it is an outlier.

You read synthesize the way you read a diff in Module 3: line by line. If a sentence makes a claim, ask what source supports it. If no source does — if it is a smooth connector sentence that sounds right but is not actually supported — that is the research equivalent of a silent bug. It will slip past every casual reader, and it will still be wrong.

Why fabrication risk is highest here CORE

Of all the directed work in this course, research is the one where the agent is most likely to produce a confident wrong answer — and the one where that wrong answer is most likely to ship without being noticed.

Three things stack.

The model wants to sound helpful. Language models are trained, among other things, to produce answers that look complete. When the real answer is “I could not find a reliable source on this,” the model’s strong pull is to produce a plausible-looking citation instead. That is where hallucinated citations come from — references to papers that do not exist, quotations that were never written, URLs that 404. Research agents with retrieval tools hallucinate less than raw chat, but they still hallucinate. Sometimes the agent retrieves real pages and then invents a quote that is not in any of them.

Research outputs are low-feedback. If a coding agent writes broken code, the test fails and you know. If a research agent writes a confident sentence supported by a fabricated source, no alarm goes off. The only way to catch it is to open the source. Nothing else will catch it. This is the exact reason Module 4 installs you do not cite what you have not opened as its operational safety rule.

The reader trusts the format. A research brief with a list of footnotes and URLs looks authoritative. It inherits the visual credibility of a genre built by humans who actually checked. A reader — including you, the student who wrote the prompt — will tend to trust that format unless they have a reason not to. That is exactly the trust the fabricated citation is borrowing.

The defense is not cleverness. The defense is a habit: every source in your final output has been opened, skimmed, and confirmed by you. You are the only part of the system that can reliably catch this. That is the habit this module is here to install.

When a research agent is the right tool, and when it is not CORE

Not every information request needs a research agent. Three categories help.

Well-suited to a research agent. Questions where the answer depends on sources outside the model’s head, where you want a range of viewpoints or sources, and where “I need to be able to defend this” matters. Examples: “What do current state-by-state laws say about X?” “How have three major news outlets covered Y this month?” “What are the arguments, pro and con, for switching our small business from accounting software A to accounting software B?” “Summarize what the last three years of published research says about sleep and high-school academic performance.”

Well-suited to a summary (no retrieval). Questions where you want a concept explained, a term defined, an idea mapped out, and the model’s training is plenty. Examples: “Explain what a Fourier transform is.” “Give me the gist of the Cold War in one page.” “What does ‘ontology’ mean?” Using a research agent here is not wrong — it will still work — but you are paying for retrieval you did not need.

Poorly suited to either. Questions where the right move is to read the primary document yourself. Examples: “What does this 2,000-word essay argue?” (read the essay). “What does clause 4B of this contract mean?” (read the contract, possibly with an agent helping you parse, but with your eyes on the text). “What does my grandmother mean in her letter?” (read the letter). The research agent can help you index long reading, but it cannot substitute for you engaging with a single document that matters.

There is also a fourth category worth naming: questions that are dangerous to let an agent answer alone, because the stakes of a confident wrong answer are high and the verification cost is also high. Medical questions about your own body, legal questions about your own case, financial decisions you are about to act on — these are not “research agent says” questions. They are “research agent helps me find the right specialist” questions.

Picking a topic to thread through the module CORE

This is the first module where your homework is not a standardized exercise. It is a real research question you actually need answered. The rest of Module 4’s practice is built on this topic — you will scope it in 4.2, fact-check a claim inside it in 4.3, synthesize it in 4.4, and ship three outputs about it in 4.5. So pick a topic you actually care about, or at least actually need.

Three decent sources for a topic:

A school paper or assignment. If you are writing a paper this term on a real topic, use that. The assignment’s existing scope is a head start, and you end the module with work your teacher can grade.

A decision in your life. A summer job you are considering, a college you are weighing, a product you are about to buy, a software choice for a side project, a skill you are considering investing a year in. Real decisions make good research topics because you actually care whether the answer is right.

A question inside the work a parent does. A homeschool parent running a small tutoring business, a family business, a freelance practice — there is almost always a bounded research question inside that work they would pay to have answered honestly (“which of these three invoicing tools is actually right for us?”). Ask if you can pick theirs.

What makes a good topic: it is bounded (you can describe the answer in a paragraph, not a book), the answer depends on sources outside the model, and you will actually care about the quality of the answer.

What makes a bad topic: it is too broad (“tell me about AI”), it is purely conceptual (“what is consciousness”), it is a summary disguised as research (“explain the French Revolution”), or you picked it because it sounded impressive and you do not actually need the answer.

One topic. Commit to it in the project checkpoint at the end of this lesson. You can change it once, in Lesson 4.2, if scoping reveals it does not work. After 4.2, you are locked in.