Ask the wrong kind of research agent and you'll spend longer checking its report than the research itself would've taken. Match the agent to where your answer lives: report writers for the open web, literature agents for published science, and search APIs if you're a developer building research into your own product.
You've got a question that would normally eat an afternoon: run the searches, read what comes back, turn it into something you can use. That legwork is what an AI research agent promises to take off your hands, but the label now covers very different products. The nine below are for analysts, students, founders and developers who want multi-step research done for them, grouped by the kind of research each one does.
The short answer for most researchers
If your question is about the world at large, start with a tool that writes a report from the open web. If it's about what published studies say, use one built for scientific literature, because it'll cite papers rather than blog posts. And if you're a developer building research into your own app, you don't want a finished agent at all; you want a search API sitting underneath one.
What I wouldn't do is take one tool and ask it everything. These agents are strong at the job they were built for and noticeably weaker outside it, and the report reads just as confidently either way. That confidence is the thing to watch, because a smooth paragraph with a citation attached looks finished whether or not the source says what the paragraph claims.
Three products sharing one name
Here's what the marketing pages won't spell out side by side. Only some of these are finished research agents. Perplexity and Gemini both deliver a written report from a question. Elicit, Undermind and Paperguide stay inside scientific literature. Exa and Tavily describe themselves as search APIs for building your own agents, so there's no report waiting at the end unless you write the code that produces one.
Gemini's page adds one more detail worth weighing: its Deep Research can draw on Gmail and Drive as well as the web. For some people that's the whole appeal, because their best sources are their own documents and emails. For others it's a reason to check exactly what the tool can see before switching that access on.
So test first for the thing that'll hurt you most if it goes wrong. For a report writer, that's whether its sources say what the report claims. For a literature agent, it's whether it found the papers you already know matter in your field. For a search API, it's whether the results come back fresh and clean enough to build a product on.
Where each agent does its research
Reports from the open web
The two deep research modes from Perplexity and Google sit here, alongside Manus, which goes a step further than research and talks about executing tasks and automating workflows. That's handy if the research is only the first half of a job, and more than you need if all you want is an answer you can read and check.
Published science
Three tools in this list read academic papers, and Asta joins them from Ai2 as a wider set of agents and benchmarks for scientific research. Students writing a literature review and researchers checking the evidence on a question will spend most of their time in this group, and they'll care far more about which papers were found than about how polished the summary sounds.
Parts for building your own
The two search APIs belong to developers. They return web data that your own agent can reason over, which gives you control over the steps and the sources, and leaves you to build everything a finished product would've handed you on a plate.
Nine agents at a glance
| Tool | Kind of agent | Research it suits |
|---|---|---|
| Asta | Scientific research agents | Scientific research with agents and benchmarks |
| Elicit | Scientific literature agent | Synthesising scientific evidence |
| Exa | Search API for agents | Developers feeding web data to their own agents |
| Gemini Deep Research | Web and workspace research | Web reports that can also use Gmail and Drive |
| Manus | General-purpose agent | Tasks that go past answering a question |
| Paperguide | Literature review platform | Literature reviews and systematic reviews |
| Perplexity Deep Research | Web deep research | Written reports on open web questions |
| Tavily | Search API for agents | Connecting AI agents to the live web |
| Undermind | Scientific literature agent | Finding papers by reading full texts |
Nothing in the table is ranked. It's there so you can find the two or three that match your kind of research and ignore the rest.
Every agent, alphabetically
In no particular order beyond the alphabet. Each line sums up what the maker says its own tool does.
Asta
Best for: Scientific research work, through what Ai2 calls an agentic ecosystem of AI agents and benchmarks for advancing discovery.
Elicit
Best for: Researchers synthesising scientific evidence with a research agent and report generator; the vendor's pitch is helping researchers be more evidence-based.
Exa
Best for: Developers who want, in the vendor's words, one endpoint for web data, deep research and structured content inside their own AI agents.
Gemini Deep Research
Best for: People already in Gemini, since this agentic feature browses websites and can draw on Gmail, Drive and Chat; the vendor calls it a personal research assistant.
Manus
Best for: Work that needs doing rather than just answering, with the vendor describing an action engine that executes tasks and automates workflows.
Paperguide
Best for: Academics running literature reviews and systematic reviews with research agents and reference management; the vendor promises auditable systematic reviews.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Perplexity Deep Research
Best for: Analysts who want a full written report: it runs many searches, reads sources and works through the material, per the vendor.
Tavily
Best for: Builders connecting AI agents to the web through search, extract, crawl and research APIs, which the vendor sums up as one secure API.
Undermind
Best for: Scientific literature searches that read full texts, not just abstracts, and cite every source, according to the vendor.
Test them on a question you can already answer
Don't judge a research agent on its demo question. Choose something you already know the answer to, ideally a topic where you've read the main sources yourself, and give the same wording to two agents from the same group. That way you're grading the tool, not learning the subject at the same time.
Then check three things. Did it find the sources you'd expect, or miss an obvious one? When it states something, does the cited source say that if you open it? And how long did it take you to check the result compared with doing the research yourself? A report that takes an hour to verify hasn't saved you much.
If you're a developer comparing the APIs, the equivalent test is one query your product will run every day, sent to both, with the raw results laid side by side and read line by line.
Checking advice before you repeat a report
Treat every report as a first draft from a quick, well-read assistant who's never met your field. It'll get you to the right sources faster than you'd get there alone, and now and then it'll state something its own citation doesn't support. Open the sources for any claim you're going to repeat in a meeting, a paper or a pitch.
If you'd like a wider set of tools that help with reading and note-taking rather than running whole investigations, my list of AI research assistants covers that ground. For a walk-through of one of the big report writers, see my guide to Google's deep research mode, and if ChatGPT is already open in another tab, here's how to use ChatGPT's deep research feature.
Do you make an AI research agent or research API? Tool makers can ask to be considered for a featured place on this list.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.
Frequently asked questions
What's the difference between an AI research agent and a chatbot?
A chatbot answers from what it already knows or from a quick search. A research agent runs several searches in a row, reads what it finds and writes up the result, so it takes longer and should cite its sources.
Can an AI research agent replace a literature review?
It can speed up the search and the first summary, but you still need to read the key papers and decide what counts as evidence. Use it to find and sort studies, not to judge them.
Are AI research agents accurate?
They're often right and sometimes confidently wrong. Check any claim you plan to use against the source it cites, and be most careful with recent events and narrow topics.
Can a research agent read my own documents?
Some can. One of the web report tools here says it can draw on email and cloud files as well as the web, so check what access you're granting before you switch that on.
Do I need to code to use a research agent?
Not for the report writers or the literature agents. The search APIs are for developers who want to build research into their own software.
Are any AI research agents free?
Several say they offer a free plan, often with limited use. Check each tool's own site, since what's included can change.
