Undermind vs Agent Bayes: one finds the papers, the other shows the page
In its own test, Undermind found 73% of the relevant papers in 10 minutes. Agent Bayes cites each claim to the passage and the page number printed on the article, then checks your wording against it. What each one is for, task by task.

In Undermind's own test, its deep search found 73% of the relevant papers in 10 minutes, while general AI agents running their own web searches found 29% to 39%.
Undermind is built to find the papers you missed. Agent Bayes is built to show the page that supports each sentence you write from them. Both read full text, both connect to Claude Code and Codex, and both keep your work between sessions, so the useful question is which problem you have right now.
| Your task | Better fit | Why |
|---|---|---|
| Find papers on a question you are new to | Undermind | Searches 200M+ papers, follows citation trails, stops when nothing new appears |
| Keep a saved search up to date | Undermind | Adds newly published papers to past searches |
| Ask questions across full texts, including figures | Both | Each one reads whole PDFs |
| Cite a claim to the exact passage and page | Agent Bayes | Each citation stores the passage and the printed page number |
| Check whether your wording overstates the source | Agent Bayes | Scores your sentence against its cited passages, suggests tighter wording |
| Review your own draft claim by claim | Agent Bayes | Lists every claim in a draft and checks each one |
| Develop an argument over months | Agent Bayes | A mindmap of cited claims, with saved versions |
| Work from your Zotero PDFs | Agent Bayes | A Zotero plugin sends your PDFs and their details |
| Use it from Claude Code or Codex | Both | Each one connects as a tool |
If someone in your lab is starting a literature review, send them this table: search in one tool, cite from the other.
The capabilities below come from Undermind's public pages as checked on November 4, 2026, and from the current Agent Bayes code. We build Agent Bayes.
Undermind searches until new searches stop finding papers
A keyword search misses papers that use different words for the same idea, and it never tells you what it missed. You run five queries, skim the top results, and move on. Undermind was designed around that gap.
Its page on connecting AI assistants describes a deep search in stages: a search by meaning over more than 200 million papers, then citation trails, then the work of important authors. The search ends when a pass adds nothing new, after about 2.5 minutes on average. The result is a ranked list of papers with a written summary.
The stopping rule is the most useful design choice. Undermind's 2024 white paper described how the rate of finding relevant papers falls off exponentially as a search goes on. Tracking that curve let the system estimate when it had found nearly all of the relevant work.
Undermind measures its search with recall. On its homepage and enterprise page, Undermind gave each system 10 minutes per question on 23 research goals, with relevance judged from abstracts by a GPT model.
Undermind reached 73% recall. General AI agents running their own web searches reached 29% to 39%. The test did not include Agent Bayes, which searches only the libraries you choose.
Undermind also adds newly published papers to past searches. Finding the right papers comes first in a new project, and Undermind is built for that first step.
Both tools read the full text of a paper
A common objection to AI search tools was that they read only abstracts. That objection does not separate these two products.
Undermind's homepage says it fetches open-access PDFs and lets you upload the rest, and that its agent can read them all. Its agent answers questions across many papers at once, including figures, tables, and equations.
Agent Bayes reads the PDFs you add to a library, a collection of papers you choose. Text recognition keeps columns, tables, captions, and footnotes, and each passage gets English key points, whatever the language of the paper. Figures get a written description, and when you chat with a paper or a library, the agent can open the page image to check a detail.
So both tools read papers in full. The difference shows up one step later, when you copy a sentence and its citation into your chapter.
The difference is what a citation points at
A citation can point at three different things: a paper, a passage, or a page you can put in a bibliography. Each level answers a different question: which paper, where in it, and which page to cite.
Undermind describes its reports at the level of papers. Its agent "creates and edits Markdown notes, syntheses, and reports in the workspace," and their citations "link back to the source papers." None of Undermind's public pages that we checked describe a citation that opens a passage or names a page.
A University of Victoria library guide, written for an earlier version of Undermind, shows the extra step. It asks students to click the citation link after a claim, then look for where the source supports the claim in the journal article.
A link to a paper tells you where to start reading. It does not tell you which sentence of that paper supports your claim.
A page-level citation removes that search. In Agent Bayes, a citation stores the passage the agent used, enough surrounding text to read it alone, and its page range. Click the citation, and the built-in PDF reader opens that page.
The page number is the printed one. A journal article might start on page 412 while the PDF calls the same page 1, and publishers sometimes add a cover sheet that shifts every page again. Agent Bayes keeps both numbers, as we explain in our post on journal page numbers and PDF positions.
A real paper can still be the wrong source
Undermind's homepage says, "Every citation is verified," and its public pages do not describe what that verification compares. Agent Bayes runs a specific check: whether the cited passages support the sentence as you wrote it.
A real paper can be the wrong source for the sentence that cites it. The paper found the effect in one population and your sentence says all of them. Or the sentence repeats a position the authors reported and then rejected. Our guide to checking whether a citation supports a claim walks through these cases.
In Agent Bayes this check is called AI Verification. Select claims in a mindmap, the diagram of cited claims you build with the agent, and press F3. The agent reads the passages cited for each claim and returns a support score from 0 to 100, the phrases the passages do not support, and four tighter phrasings. It costs about 1 credit per claim.
The same check runs on a whole draft. A paper review, called AI Peer Review Paper in the app, lists every claim in a paper from your libraries, including your own draft, then checks each claim against the other papers. Each claim gets a verdict: Supported, Partly supported, Contested, Contradicted, or Not found in your library.
On a two-page test draft with seven planted mistakes, a full review caught all seven in under 2 minutes, for 5.5 credits.
A report ends a search, a mindmap keeps the argument
Undermind's unit of work is a search and the report written from it. Reports stay in a workspace with folders, and you can keep editing them. That suits a briefing, a reading list, or a scoping memo.
A thesis chapter is a different kind of document. It is a set of claims that changes as you read, and each claim needs its own evidence. In Agent Bayes, a mindmap holds those claims, each with its own citations, and both you and the agent edit them.
Three properties follow:
- When two of your sources disagree, both positions stay on the mindmap as separate claims with separate evidence. We explain why that matters in preserving scholarly disagreement.
- Every agent edit is recorded. You can replay what changed, undo it, and save a named version of a state you trust. Our post on version control for AI research edits covers the design.
- The agent searches only the libraries attached to your project. When someone asks why a claim rests on these papers, the answer is that you chose them.
Use Undermind to find papers and Agent Bayes to cite them
The two tools cover different steps, so many researchers will want both. Undermind answers "what should I read," and Agent Bayes answers "where does it say that."
A practical workflow for a literature review:
- Run a deep search in Undermind on your question and read the ranked list.
- Save the papers you keep in Zotero, with their PDFs.
- Send those PDFs to an Agent Bayes library with the Zotero plugin.
- Ask Agent Bayes to build cited claims from them, then verify the claims you will write into your chapter.
Agent Bayes has also been measured on answering questions from a set of papers. On three public benchmarks that do that, the strongest result came on IRPAPERS, 180 questions over 166 information-retrieval papers. The benchmark's model judge accepted 97% of the Agent Bayes answers, against a best published result of 82%.
On LitQA2, 67 detailed biology questions, it answered 61 correctly (91%) and gave no wrong answers. On five of the other six, it said the papers did not settle the question.
Try it on five papers you already know
Install the Zotero plugin, open its Agent Bayes panel, and send five papers you know well to a new library. Indexing costs about 0.35 credits per page, so a 20-page article costs around 7 credits, as the credits guide explains.
Attach the library to a project and ask the mindmap chat: "Compare these papers' findings and add supported claims with citations." On the lowest-cost model that run costs a few credits. Click a citation and check that the reader opens the page you expected.
Then select a claim and press F3 for AI Verification, about 1 credit per claim on the Starter plan or higher. The score shows how well the cited passages support your wording.
New posts, straight to your inbox
No newsletter fluff, just an email when we publish something new.
Email me when a new post is published on the Agent Bayes blog. You can unsubscribe anytime. We'll first send a confirmation email, and we only use your details for this. See our Privacy Policy.