ProjekinAja. Agentic AI Stage 3 · Working With Agents Lesson 3.9
● Stage 3 · Lesson 3.9

Agentic Research & Writing

Beyond code, agents shine at knowledge work — gathering, synthesizing, and drafting. But research agents carry a special danger: confident, well-cited nonsense.

⏱ ~14 min read 🔎 Knowledge work 🧪 Verify everything

🎯 By the end you'll understand

What a research/writing agent actually does (decompose → search → extract → cross-check → synthesize → draft), why it's so powerful, the unique failure mode of fabricated citations, how to brief and constrain it, and why your real job is editor and fact-checker — never stenographer.

KEY INSIGHTThe one idea to take away

A research agent doesn't give you truth. It gives you a fast, fluent, footnoted draft — and footnotes are not the same thing as facts.

Coding agents have a brutal honesty: the code runs or it doesn't (Lesson 3.8). Research agents have no such referee. A made-up statistic looks exactly like a real one, and a fabricated citation looks more trustworthy than no citation at all. That polish is the trap. The skill you're building this lesson is the discipline to treat a beautiful research report as a claim to be checked, not an answer to be trusted.

🔑 Hold onto this

Citations are a verification surface, not a guarantee. A good research agent makes its claims checkable by pointing at sources. Your job is to actually follow those pointers for anything that matters — because the agent can, and sometimes will, cite a real-looking source that says no such thing.

CORE CONCEPTWhat a research agent actually does

"Research agent" sounds mysterious, but the pipeline is concrete. It's an agent loop (Lesson 2.1) pointed at a question instead of a task, and it almost always runs through six stages:

1

🧩 Decompose

It breaks your big question into smaller, answerable sub-queries — exactly the task-decomposition skill from Lesson 2.5. "Is X a good market?" becomes "market size?", "growth rate?", "main players?", "key risks?"

2

🔎 Search

It fires those sub-queries at sources — web search, internal docs, databases — often several in parallel rather than one at a time.

3

📄 Read & extract

It opens the results, reads them, and pulls out the specific facts, numbers, and quotes that bear on each sub-query.

4

⚖️ Cross-check

It compares what different sources say, notices agreement and disagreement, and (ideally) flags where they conflict instead of silently picking one.

5

🧠 Synthesize

It weaves the extracted findings into a coherent answer to your original question — not a list of search results, but an argument.

6

✍️ Draft with citations

It writes the report in prose, attaching a citation to each claim so you can trace where it came from.

The serious "deep research" features you've seen in modern tools are usually a multi-agent system (Lesson 2.7): a swarm of searcher agents chasing different sub-questions in parallel, a synthesizer that pulls their findings together, and sometimes a critic that pokes holes before the draft reaches you. That parallelism is exactly why these runs can read a hundred pages in the time it takes you to read one.

📖 Research agent

An agent that answers a question by decomposing it into sub-queries, searching multiple sources, extracting and cross-checking findings, and synthesizing a cited draft — the knowledge-work cousin of the coding agent.

PRACTICALWhy it's genuinely powerful

Once you see the pipeline, the strengths are obvious — and real. This is one of the highest-leverage uses of agents today:

Notice what's not on that list: being right. Breadth, speed, and stamina are about coverage and fluency, not accuracy. The agent is a phenomenal research assistant and a dangerous research authority. Keep those two roles separate in your head.

SEE ITWatch a research run — then verify it

Here's a real-shaped research run for one question. Press Run the pipeline to watch it decompose, search, extract, cross-check, synthesize, and draft. Then — and this is the point — press Verify the citations. One of those clean-looking sources is fabricated.

🔎 Interactive · Research Agent Pipeline

Question: "Is the home espresso machine market a good one to enter in 2026?"

Six stages, played in order — deterministic, hand-authored.
1

🧩 Decompose

Sub-queries: market size · growth rate · top players · buyer motivation · key risks.

2

🔎 Search (parallel)

5 searchers dispatched → 14 sources returned across the sub-queries.

3

📄 Extract

Pulled: a market-size figure, a CAGR, 3 named brands, 1 buyer-trend stat, 2 risk notes.

4

⚖️ Cross-check

Two sources agree on growth; one outlier disagrees → noted, not hidden.

5

🧠 Synthesize

Findings woven into a verdict: attractive-but-crowded.

6

✍️ Draft with citations

Report written; each claim tagged to a source below.

Illustrative numbers, hand-authored to teach the mechanism — not a real market report.

WEAKNESSThe special risks of research agents

A coding agent's bug shows up as a stack trace. A research agent's bug shows up as a confident, well-formatted sentence with a footnote. There's no compiler to catch it. Four failure modes deserve your suspicion:

1 · Hallucinated facts. The model can invent a statistic — a market size, a percentage, a date — that sounds plausible and is simply wrong. This is the familiar hallucination problem, now wearing a research report's clothes.

✗ Hallucinated citations — the nasty one

Worse than a fake fact is a fabricated source: a real-looking citation — plausible author, plausible title, plausible year — attached to a claim the source never made. Sometimes the source doesn't exist at all; sometimes it exists but says something different. The footnote makes the lie look more credible. This is why you verify the link, not just the sentence.

2 · Source quality. The agent can cite junk — a random blog, a marketing page, a forum guess — with the same confident tone it uses for a peer-reviewed study. It doesn't natively rank a press release below a regulator's filing unless you tell it to.

3 · Confirmation bias. If your prompt leans ("show me why this market is great"), the agent will helpfully find evidence that says so and quietly under-weight what contradicts it. It tends to find what it expects to find — which means a biased question yields a biased, well-cited answer.

FRAMEWORKHow to drive a research agent well

The same levers that make any agent reliable apply here, sharpened for knowledge work. Four moves carry most of the weight:

🎯

Ask a sharp question

Vague in, vague out. "Tell me about espresso machines" wanders; "What's the 2026 market size, growth rate, top 3 players, and main barrier to entry for home espresso machines?" focuses the whole pipeline. This is the briefing skill from Lesson 3.2 — and ask neutrally, so you don't pre-load the bias.

🚧

Constrain the sources

Tell it what to trust and what to avoid: "Use industry reports and official filings; ignore vendor marketing and forums." Source constraints are how you fight the junk-citation problem before it happens.

🔗

Demand checkable citations

Require a source per claim, with enough detail to find it. A claim you can't trace is a claim you can't trust — treat it as unsupported.

🧪

Verify what matters

For anything important, actually open the cited source and confirm it says what the report claims. This is verification (Lesson 3.6) applied to knowledge work — you scale your checking to the stakes.

You don't verify every sentence — that would defeat the speed-up. You verify the load-bearing ones: the numbers you'll act on, the claims you'll repeat, the facts a decision hangs on. The agent's gift is making those claims easy to check; spend the saved time checking them.

PRACTICALDrafting and writing — taste is your job

Research blurs into writing, and the same principles carry over. Agents are excellent at the mechanical heavy lifting of prose:

What they cannot supply is taste, judgment, and accountability. The model doesn't know which point actually matters to your reader, whether the tone fits, or whether a claim is one you're willing to stand behind. Those are yours. A draft is raw material; the editing — choosing what to keep, what's true, what's worth saying — is where your value lives.

Your role with a research or writing agent is editor and fact-checker, not stenographer. You don't transcribe its output — you interrogate it.

If you find yourself copy-pasting a research report into a decision deck without reading the sources, you've slipped from editor to stenographer — and you've inherited every fabricated citation in it, with your name on the cover.

FRAMEWORKA trust checklist for research output

Before you act on or forward an agent's research, run it through this. The more "no"s, the more skeptical you should be:

Ask of the report…Why it matters
Does every important claim have a citation?An uncited claim is the model's guess — treat it as unverified.
Did I open the cited sources for the key claims?Catches fabricated citations and "source says something else."
Are the sources actually credible?A confident tone can't upgrade a forum post into evidence.
Did it show disagreement, or suspiciously one-sided?Real research has conflict; a tidy, one-sided answer hints at confirmation bias.
Was my question neutral, or did it lean?A leading prompt produces a leaning, well-cited answer.
Would I stake a decision on this as written?If not, you've found exactly what still needs checking.
💡 The market-summary habit

Take the espresso run above. The "$4.6B" and "9% growth" feel authoritative — but before they go in a business case, you open [A] and [B] and confirm the figures. When you do, you discover the survey behind the "62% of buyers" stat [D] doesn't exist. The report was 90% useful and 10% fabricated — and only verification told you which 10%.

WEAKNESSCommon mistakes to avoid

✗ "If it has citations, it must be true."

Citations make a claim checkable, not correct. A fabricated source is more dangerous than no source, because it borrows the look of rigor. Follow the footnote before you believe it.

✗ "The report looks thorough, so I'll skip checking the sources."

Polish is the agent's specialty and tells you nothing about accuracy. Length and formatting are not evidence. Open the sources for the claims that matter — every time the stakes are real.

✗ "It's well-written, so I'll treat the draft as final and authoritative."

A draft is a starting point you edit and fact-check, not a verdict you forward. Treating it as authoritative is how confident, well-cited nonsense ends up in your decisions — with your name on it.

KEY INSIGHTWhere this connects in your course

  • Task decomposition (2.5) — breaking your question into sub-queries is exactly the decomposition skill, applied to research.
  • Multi-agent systems (2.7) — "deep research" is usually a swarm: parallel searchers, a synthesizer, sometimes a critic.
  • Context & tools (2.4) — search, fetch, and read are the tools that let the agent reach beyond its training and into live sources.
  • Briefing (3.2) — a sharp, neutral question with source constraints is what aims the whole pipeline.
  • Verification (3.6) — checking claims against their sources is verification applied to knowledge work; scale it to the stakes.
  • The "deep research" idea — the productized version of this pipeline you'll meet in real tools is just these six stages, run by a multi-agent team at scale.

Next up: agents that don't just answer or draft, but run on their own on a schedule — Lesson 3.10 · Agentic Automation.

GLOSSARYKey terms introduced

Research agent
An agent that answers a question by decomposing it, searching sources, extracting and cross-checking findings, and synthesizing a cited draft.
Synthesis
Weaving findings from many sources into one coherent answer to the original question — an argument, not a list of results.
Sub-query
A smaller, directly answerable question the agent splits a big question into (e.g. "market size?", "growth rate?").
Citation
A pointer attaching a claim to its source so you can trace and verify it. A verification surface, not a guarantee of truth.
Fabricated source
A real-looking citation — plausible author, title, year — for a claim the source never made, or that doesn't exist at all.
Cross-checking
Comparing what multiple sources say to surface agreement and (importantly) disagreement, rather than silently picking one.
Confirmation bias
The tendency to find evidence that confirms what was expected; a leading prompt yields a leaning, well-cited answer.

ACTIVE RECALLCheck your understanding

Answer in your own words first, then reveal.

1Name the six stages of a research-agent pipeline.›

Decompose (into sub-queries) → search (often parallel sources) → extract (read & pull facts) → cross-check (compare sources) → synthesize (weave into one answer) → draft with citations. It's a multi-agent loop pointed at a question.

2Why is a fabricated citation more dangerous than a hallucinated fact alone?›

Because the footnote borrows the appearance of rigor. A real-looking source attached to a false claim makes the claim look more credible, lowering your guard. The only defense is following the citation to confirm the source actually says it.

3What four moves make a research agent reliable?›

Ask a sharp, neutral question (3.2); constrain the sources it may use; demand checkable citations per claim; and verify the load-bearing claims against their sources (3.6). Scale verification to the stakes — check what a decision hangs on.

4What's your real role with a research or writing agent?›

Editor and fact-checker — not stenographer. The agent supplies breadth, speed, and a fluent draft; you supply taste, judgment, source-checking, and final accountability. A draft is raw material you interrogate, not a verdict you forward.

Fast, fluent, footnoted — now go check the footnotes. Next: automation →