Cursor AI Agents Solve a Research-Level Math Challenge After Running Autonomously for 4 Days

Cursor claims its autonomous AI system has produced a novel solution to a research-level mathematics problem after running autonomously for four days.

The company said the system tackled problem Six of the First Proof challenge, a research-level mathematics benchmark comprising 10 previously unpublished problems contributed by mathematicians and designed to test whether AI can produce original proofs.

According to Cursor co-founder Michael Truell, the agents generated a proof that may outperform the official human-written solution. The system ran continuously for four days, exploring approaches autonomously without prompts or hints.

“We believe Cursor discovered a novel solution to Problem Six of the First Proof challenge, a set of math research problems that approximate the work of Stanford, MIT, Berkeley academics,” Truell wrote on X. “Cursor's solution yields stronger results than the official, human-written solution.”

The result has not yet been academically verified. However, Cursor says early feedback from experts suggests the proof may be valid.

“We're still waiting for final expert review from Nikhil Srivastava (Associate Professor of Mathematics, UC Berkeley) or Daniel Spielman (Sterling Professor of Computer Science, Yale University), but we have received feedback from spectral graph theory expert Yang Liu that our proof is likely correct,” Truell said, adding that Stanford mathematician Jan Vondrák also reviewed the solution and found it appeared correct to the best of his knowledge.

The proof reportedly uses the Marcus–Spielman–Srivastava interlacing polynomial method, a technique from spectral graph theory. Truell said the approach differs from that used in the existing solution and yields stronger guarantees. “It used the Marcus-Spielman-Srivastava interlacing polynomial method, a different approach from existing solutions,” he wrote. 

“Two concrete improvements: the constant c goes from 0.03 to 0.13, and the solution partitions the entire vertex set into light components rather than just a subset.”

Cursor saidthe experiment hints that systems originally designed for large-scale software engineering could generalise to research problems beyond coding.

The company ran the experiment using the same multi-agent harness it recently described in a research post about autonomous coding. The system coordinates hundreds of AI agents through a structured workflow that separates planning and execution roles.

In that setup, planner agents generate and refine tasks while worker agents execute them independently. The architecture was previously used in large-scale experiments in which agents collaborated for weeks to write millions of lines of code, including building a web browser from scratch.