Cubicles
Cubicles

How to use an AI agent for web research you can actually trust

Agents are fast, tireless readers and confident guessers. A good research setup keeps the first and catches the second.

Blog ·

Research is where AI agents save the most time per task. Comparing twenty suppliers, tracking a dozen competitors' prices or collecting every relevant regulation takes a person days of tab-switching. An agent reads all of it in an hour.

The catch is trust. A model that can't find a number may write a plausible one, and a confident summary is easy to believe. This guide covers how to set up research you can rely on.

Deep research tools vs an agent with a computer

ChatGPT, Claude, Gemini and others offer a “deep research” mode that searches the web and writes a report. For a one-off question, that's often all you need. An agent running on its own computer is different:

Deep research in a chat appAn agent on its own computer
Best forOne-off questionsOngoing and repeated research
SourcesPublic web pagesPublic pages, plus sites and databases you sign it in to
OutputA report in the chatFiles you keep: spreadsheets, Markdown, CSV
Runs on a scheduleNoYes: daily, weekly, whenever you set
Remembers last timeNot usuallyYes, from the files it saved
Can act on findingsNoYes, such as filling in a form or messaging you

Use deep research for a question you'll ask once. Use an agent when the research repeats, needs logins, or should end in files and actions rather than a chat reply.

What research agents are good at

  • Breadth. Reading fifty pages to find the ten that matter.
  • Structured extraction. Turning messy pages into a table with the same columns for every row: price, minimum order, lead time, contact.
  • Comparisons. Lining up products, suppliers or plans side by side against your criteria.
  • Watching for change. Checking the same sources every week and reporting only what's new.
  • First drafts. A starting list of sources, competitors or questions that you then refine.

Where research agents go wrong

  • Invented details. A price, date or quote that isn't on any page. This is the big one.
  • Stale pages. A 2023 price list read as current.
  • Low-quality sources. SEO filler and affiliate lists ranked above the primary source.
  • Missing what's behind a login or loaded by an app, then reporting “not published”.
  • Over-confident summaries that smooth over disagreement between sources.
  • Blocked sites. Some sites challenge automated visitors, so pages can come back empty.

Write a research brief, not a question

“Find me packaging suppliers” gets a vague list. A brief gets a usable table. A good brief covers:

  1. The decision it supports. “I'm choosing a supplier for 5,000 boxes a quarter.”
  2. The scope. How many results, which regions, which time period.
  3. The fields to collect, the same for every row.
  4. Which sources to prefer, like the company's own site over directories, and which to ignore.
  5. What to do when it can't verify something: mark it, don't guess.
  6. The output: file name, format and where to save it.

“Compare five suppliers for recycled packaging that ship to Germany. For each, record minimum order, lead time, price per 1,000 units where published, certifications and a contact link. Use each company's own website first. Put a source link next to every figure. If you can't find or verify something, write ‘not found’ instead of estimating. Save the comparison as supplier-shortlist.md and list anything you couldn't check at the end.”

How to check the work

  • One source per fact. Ask for a link next to every figure, not a list of sources at the bottom.
  • “Not found” is allowed. Say explicitly that a blank is better than a guess. Models guess less when told it's fine not to know.
  • A verification pass. Ask the agent to open every link in its own table and confirm the figure appears on that page. It catches a surprising number of errors.
  • Spot-check three rows yourself, picked at random.
  • A gaps section. Ask it to list what it couldn't find and where sources disagreed.
  • Dates on everything. Ask for the date each page was published or updated, where shown.

Recurring research: the real payoff

One report saves you a day. The same report every week, done without you, saves a day every week. Good recurring jobs:

  • Competitor pricing, every Monday, with changes since last week highlighted.
  • New tenders, grants or job postings that match your criteria.
  • Regulation and policy updates from the official sites in your industry.
  • Supplier stock and lead times before you place an order.
  • Mentions of your company or product in the press and forums.

Ask the agent to save each run as a dated file, compare it with the last one, and message you only what changed. That keeps the summary short and gives you a history. Our guide on how to schedule an AI agent covers the setup.

Which agent for research?

Hermes Agent suits research that repeats with small variations, because it keeps memory and refines its own skills as it goes. OpenClaw suits research you want delivered to your chat apps on a schedule. Both work well; see OpenClaw vs Hermes Agent. Whichever you choose, a model that's reliable at tool calls and honest about gaps matters more. See choosing a model for your agent.

Set up a research agent on Cubicles

A research agent needs a real browser, a place to save files and a computer that stays on. On Cubicles, each computer is a full Ubuntu desktop with Chrome, Firefox and LibreOffice for spreadsheets, and Hermes and OpenClaw come installed.

  1. Sign up. Your first computer is free.
  2. Choose New agent, pick Hermes Agent or OpenClaw, add your model and key, and pick a chat app.
  3. Choose Continue to its computer and create it. The agent is set up there a few minutes later.
  4. Send it your first brief, and open the computer in your browser to watch it read.
  5. When the first report looks right, ask it to repeat the job on a schedule.

If a site asks for a sign-in or a human check, you can take over the screen, deal with it, and hand back. Our research workspace guide has a ready-made first task.

Frequently asked questions

Can an AI agent do market research?

Yes, especially the gathering part: finding sources, extracting prices and details into a table, and tracking changes over time. The judgement on what it means is still yours.

Is AI research accurate?

Not automatically. Models can invent details when they can't find them. Ask for a source link next to every figure, allow it to say not found, and have it verify its own links before you rely on the results.

What's the difference between deep research and a research agent?

Deep research in a chat app answers one question with a report. A research agent on its own computer can repeat the work on a schedule, use sites you sign it in to, save files and compare each run with the last.

Can a research agent read sites that need a login?

Yes, if you sign it in on its computer with an account you're allowed to use that way. Check the site's terms, and give it an account of its own rather than yours where you can.

Which AI agent is best for research?

Hermes Agent suits repeated research that should improve over time. OpenClaw suits scheduled research delivered to your chat apps. The model you connect matters as much as the agent.

Give your agent a computer