Cursor: agent swarm discovers novel solution to Problem Six of the First Proof challenge

TL;DR

Cursor CEO Michael Truell announces that Cursor’s agent swarm — the same long-running system that previously “built a full browser from scratch unsupervised” per follow-on reporting — solved Problem Six (a spectral-graph-theory problem) of the First Proof challenge, a 2026-02 math research benchmark created by a group of Fields Medalists and MacArthur Fellows to test “research-grade” rather than “competition-style” mathematical reasoning. The claim, per Truell, is that the agent produced a solution yielding stronger results than the official human-written solution, after four days of zero human nudging. Truell frames this as evidence that Cursor’s agent-coordination technique “might generalise beyond coding.” No paper, write-up, or proof artifact is linked from the tweet; this page captures the product claim as it circulated.

Key claims

Cursor’s agent swarm produced a novel solution to Problem Six of the First Proof challenge that the team believes yields stronger results than the human-written reference solution [tweet body].
The First Proof challenge is a research-math benchmark released 2026-02, designed by prominent mathematicians (including Fields Medalists and MacArthur Fellows) explicitly to move past pattern-matchable competition-style problems [external context from etn.’s summary thread, not in the original tweet].
The agent run was ~4 days with zero human nudging, on a spectral-graph-theory problem; same agent infrastructure previously reported to have “built a full browser from scratch unsupervised” [external context from the AI Investor’s commentary thread, not in the original tweet].
Truell frames the result as evidence that the agent-coordination technique may generalize beyond coding [tweet body + follow-on commentary].

Method

Not disclosed. The tweet is a product/result announcement. No paper, proof artifact, transcript, or technical write-up is linked. From follow-on commentary (which is third-party, not from Cursor) the setup is characterized as a long-running agent swarm operating without human intervention over multiple days on a single math problem; no architecture, training recipe, tool stack, or compute budget is given.

Results

One headline claim: novel solution to one problem (Problem Six, spectral graph theory) on one benchmark (First Proof challenge), reported to be stronger than the human reference solution.
No quantitative axes (no benchmark score, no comparison against other models on the same benchmark, no ablation, no inter-run variance).
No third-party verification of the math is referenced in the tweet.

Why it’s interesting

A real datapoint for the “general-purpose agent” framing that Introducing GPT-5.3-Codex articulates at the closed-frontier — that a coding agent is “becoming a general-purpose agent that can reason, build, and execute” — but pushed harder: this is the same agent stack moving from SWE-Bench-style coding to research-grade math without (per Truell) a new training run. That makes it the cleanest filed instance so far of an Agentic Software Engineering system claiming out-of-distribution transfer to a non-coding research benchmark, complementing FARS: Fully Automated Research System (FARS) and OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis which target the research-agent stack natively. It also fits the multi-day, multi-agent Inference-Time Scaling regime — the same direction as Kimi K2.5’s PARL (Reasoning RL) and MiroThinker’s interaction-depth scaling (MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling) — but at far longer wall-clock (4 days vs minutes-to-hours) on a single problem with no online human feedback. Worth tracking, but with the standard caveats for tweet-as-primary claims about math: no proof artifact, no third-party verification, no methodological detail. Swayam’s accompanying note (“maybe its worth it to just lock up a codex 5.3 with a h100 node and see if it can do some research by itself”) captures the obvious follow-up — whether the result reproduces with a different agent stack at a fraction of Cursor’s compute.