AI agents for research: 36 real examples

It created the text file as asked, but after clicking a result it took a screenshot too early and hallucinated the answer.

Operator handled tour search and lesson plan but was slow; adding three grocery items took 15 minutes and an Uber price check took five.

The draft was usable but weaker than accepted NSF summaries; thinking mode helped, though it began hallucinating and misreading the sources.

Operator collected data for about 25 of 100 schools before the author stopped it; o1 then built a working chart dashboard from that data.

Both tasks failed. The concert search returned three wrong concerts after 33 minutes, and the Anki competitor list had incorrect data.

It produced a 30-journalist list, a tiered NYC apartment list in under half an hour, and only five solid nominees out of fifty requested.

From one short prompt, Manus produced a publishable-style report with charts from 1,000 synthetic responses. The data is made up.

It returned duplicate and months-old tweets, summarized one news source, copied a single listicle, and stopped to ask permission mid-task. The author says productivity dropped.

Manus succeeded on 8 of 11 booking engines, three of those with delays, and failed on D Edge, Simple Booking and Travelanium.

The Agent worked for 14 minutes on the UNESCO task; the author tested it against o3-pro and ran four other work-style tasks.

The VICE booking took over eight minutes and four tries on the time picker, while the SE Ranking demo booking took 37 seconds.

It produced a review-ready draft with citations and a gaps synthesis in under ten minutes, versus half a day by hand.

No agent won a game. Gemini 2.5 Pro made the most progress, reaching veteran footman rank in the idle game Progress Knight.

One prompt returned a PDF deck and a USP spreadsheet in 18 minutes, with positioning, pricing models and a SWOT section.

A five-story AI news brief arrived by email from one prompt; the author says plain LLMs only draft text and cannot send it.

Perplexity Comet completed the task and stopped at the cart. ChatGPT Atlas also reached checkout but was too slow to be usable.

In 30 seconds it returned top complaints ranked by frequency, a severity versus frequency matrix, and a user journey map showing where frustration peaks.

The author came away with a health plan tailored to their genes and current health, built from the combined datasets.

The author now has two dense, searchable knowledge bases, one from YouTube videos and one from research papers, which they say speeds up understanding.

The author turned 20 videos into a browsable knowledge graph of sources and topics in their vault. A citation audit found about 60% strong matches.

In the published table, most tasks succeeded in 4 to 15 minutes, with 70% full success, 20% partial and 10% failure.

The author's weekly chart scraping and rank analysis now runs as two scheduled tasks every Tuesday, replacing manual copy-paste and taking a couple of minutes.

A full participant briefing on 40 people was ready for review within 5 minutes, replacing an evening of manual LinkedIn clicking.

In about 90 seconds it returned four posts with view and like counts, plus a suggested angle for each.

A full equity research workflow on Adyen, from data pull to investment memo, was done in under 15 minutes using prompts only.

Raghav Mehra cut morning source-checking from 45+ minutes to about 5 minutes of reading a pre-curated briefing.

The same task fell from 20 minutes in week one to 12 in week four and 8 by week six, with better competitor patterns surfaced.

It wrote a 3,000 word document with competitor pain points, 40 named target users, and an outbound playbook.

Codex returned a ranked comparison of three Belmond properties with availability and nightly rate ranges for a sample window, plus a more detailed Word doc.

Opus found no tear where the doctor reported a Grade III partial-thickness tear. The author is unsure which reading to trust.

The author cut time on literature triage and experiment setup, and says related-work sections now need less citation double-checking. No numbers are given.

It returned a price table and promotions, noted Aldi and Lidl gaps, and revised the best price per liter to Carrefour's 1,48 euro two-liter.

Research, site build and restaurant booking worked, but computer use was very slow and the call button failed.

Over 30 days it checked 3,140 coins, skipped 2,980, raised 160 alerts, and the author made 52 trades for +$9,320.

The author reports faster literature search, Zotero notes written automatically, and accurate answers about the ATS hydrology codebase through a custom ats-mcp server.

Three years of minutes (2022-2024) were extracted in about 3 to 4 hours across three sessions and loaded as one source per year.