AI Agents for the Working Mathematician: Uses and Limits
AI agents help mathematicians with literature search, conjecture exploration, and proof sketching. The mathematician verifies every step, and tools like fenbs preserve paper links and lessons from failed approaches across sessions.
Updated 5 min read
AI agents offer working mathematicians a faster way to search literature, explore conjecture directions, and sketch proof ideas. A mathematician verifies every mathematical claim independently, because agents produce plausible-sounding steps that can be wrong in subtle ways. Tools like fenbs preserve paper links, research threads, and lessons from failed approaches across sessions, so a failed line of attack this month is not repeated next month.
The gap between a plausible proof sketch and a correct proof is where mathematics lives. An agent is useful in the early stages of exploration, where generating ideas is valuable; it becomes a liability in the verification stage, where confidence without correctness is actively dangerous.
What AI agents help with in mathematical research
The most reliable applications are in information retrieval and first-draft generation: finding papers, explaining terminology in adjacent fields, and producing a starting structure for a computation or argument. Mathematicians check every substantive claim.
- Search for papers on a specific topic or result using a description of the area
- Explain terminology or standard results from an adjacent field in accessible language
- Sketch an outline of a proof approach for review and refinement
- Generate candidate examples or counterexamples to test a conjecture
- Summarize a paper's main results in plain language for a non-specialist audience
- Draft a related-work section from a list of papers and descriptions you provide
Literature search and orientation
Finding relevant work in a new area is time-consuming. An agent can help narrow the search: describe the problem, the techniques you are considering, and the results you are looking for, and ask for candidate papers. Check each suggestion on arXiv, MathSciNet, or zbMATH directly. An agent may suggest papers that do not exist or misstate results.
Orientation in adjacent fields is another good application. If your work touches functional analysis and your primary training is in combinatorics, an agent can explain standard terminology and known results at a working level. Check the claims against a textbook or a standard reference before relying on them in your own work.
When you find papers that matter, record the links in a permanent place. An agent that helped you find a key result in a session last month cannot reliably reproduce that search next month. The link belongs in your notes or in a project record where the next session can find it.
Conjecture exploration and example generation
Generating candidate examples and counterexamples is one of the more reliable uses of an AI agent in mathematics. Describe the conjecture and ask for simple cases, boundary cases, and potential counterexamples. Check each one carefully; an agent can produce outputs that look like examples but fail the conditions.
For computational exploration in specific domains, computer algebra systems like Mathematica, SageMath, or Lean provide verifiable computation. An agent can help you frame a query for one of those systems, or explain what a computation result means, but the verification belongs in the formal system.
Stop repeating failed lines of attack
In fenbs, record each significant failed approach: the conjecture, the line of attack, and why it failed. The next session reads those lessons before starting, so you do not lose time on an approach already shown not to work. Link the lesson to the relevant project task for traceability.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
Proof sketching and structure
An agent can produce a proof sketch: a high-level outline of an argument with the key steps identified. This is useful as a starting point for exploration, not as a draft to be filled in. Mathematical proofs require a level of precision that agents routinely fail at, especially in edge cases, quantifier handling, and the application of technical lemmas.
Use a sketch as a list of questions: does this step actually follow? Is this application of the lemma correct? What are the edge cases in this bound? Treating the sketch as the structure and your work as filling in details is more productive than treating the sketch as a draft proof.
Formal proof assistants like Lean 4 or Coq provide a way to check argument validity that does not depend on trusting an agent's output. If correctness is essential, the proof assistant is the right tool for verification. An agent can help you understand what a formal proof assistant is asking for or how to structure a term.
Writing and exposition
Writing papers and grant applications involves a lot of prose around the mathematics. An agent can help draft introductions, related-work sections, and plain-language summaries from descriptions and outlines you provide. Check every claim about external work against the cited paper; agents produce confident-sounding attributions that are sometimes wrong.
LaTeX drafting is a practical application: describe a mathematical structure in words and ask the agent to produce the LaTeX. Check the notation, the symbol choices, and the structure. This works well for boilerplate environments and less well for specialized notation the agent has not seen often.
How to do this with fenbs
Start with one active research project and record the key papers and failed approaches before connecting an assistant.
- Create a free account. A personal kanban board suits most mathematical research. A team kanban board supports collaborative projects with multiple researchers.
- Add the research project in fenbs. Write the problem statement, the current conjecture or goal, and the key constraints or known results that bound the space.
- Add important papers as links in fenbs with a note on why each one matters. Record each line of attack you have tried as a task, noting what it showed and why it was abandoned.
- Connect your AI app from Integrations. At the start of each research session, ask the assistant to read the project context, existing paper links, and prior lessons.
- After each session, record any new lessons, update task statuses, and save relevant paper links. Before stopping, note where the current line of exploration stands.
When AI agents are not the right tool
Do not rely on an agent to verify a proof, confirm that a result is known or unknown, or provide a correct statement of a technical lemma. These tasks require exact precision that current agents do not consistently provide. Use primary sources, textbooks, and formal systems for verification.
An agent's confidence is not correlated with its accuracy in mathematics. A step stated with certainty may be subtly wrong in a way that invalidates the entire argument. Check every step with the same rigor you would apply to a student's work, because that is roughly the reliability level you are working with.