Building an Agent Swarm with OpenAI Dots to Break a 47-Year Math Record

A Bold Community Experiment

A Reddit post has been making waves recently: someone claims to have built an "agent swarm" using OpenAI's Dots, breaking a 47-year-old mathematical record at zero cost — and formally verifying the proof using Lean. Posts like this tend to walk a fine line between "major breakthrough" and "overhyped claim," showcasing the creative potential of combining today's AI toolchains while also exposing the difficulty of verifying community content when details are sparse.

Since the source material only provides a title and a screenshot, this article will unpack the key concepts involved, helping readers understand the technical context behind this kind of experiment — while maintaining a healthy dose of skepticism about its credibility.

What Is an "Agent Swarm"?

An agent swarm refers to organizing multiple LLM-driven agents to work collaboratively, with each taking on a different role — exploration, problem-solving, verification, summarization — coordinated by some scheduling mechanism. Compared to a single model producing an answer in one shot, this swarm approach more closely resembles human team collaboration: some agents propose conjectures, others try to construct counterexamples, and others search existing literature for relevant approaches.

This architecture is particularly appealing for mathematical problem-solving. Proofs often require extensive trial and error across many branches, and a single conversation window can easily get stuck in a local line of reasoning. Multi-agent parallel search theoretically covers a much broader solution space. The post's author used OpenAI Dots as the underlying tool to build this swarm, meaning they attempted to automate and orchestrate the workflow rather than manually prompting each step.

OpenAI Dots (formally part of the OpenAI Agents SDK, sometimes used loosely to refer to its agent platform tools) provides an orchestration interface for chaining multiple model calls into a workflow. Developers can define agent roles, tool access permissions, and inter-agent communication protocols. Similar frameworks include LangGraph, AutoGen, and CrewAI — all sharing the same core idea: have each agent focus on a subtask, coordinate overall progress via message passing or shared state, and break through the length and depth limitations of a single context window. In a math setting, a typical division of labor might look like: a "proposer" agent drafting proof sketches, a "critic" agent hunting for flaws, a "translator" agent converting natural-language proofs into Lean code, and an automated compiler providing pass/fail feedback signals — forming a closed iterative loop.

Why Lean Verification Is the Critical Piece

The most noteworthy detail in the post is the author's emphasis on using Lean to verify the proof. Lean is a widely respected interactive theorem prover in the mathematics community. It requires every step of a proof to hold within a strict formal logic framework — there's no way to sneak through with something that merely "looks right."

This matters enormously. Mathematical reasoning generated by large models frequently suffers from "hallucinations" — arguments that appear rigorous on the surface but actually skip steps or contain logical gaps. If an AI-generated proof passes a full Lean check, its credibility increases significantly, because formal verification eliminates the ambiguity common in natural-language arguments. In other words, Lean verification transforms "the AI might be right" into "the proof actually holds."

That said, it's worth noting that Lean can verify "this proof is logically sound" — but it cannot determine "whether this result truly breaks a record." Establishing that still depends on accurately surveying and comparing existing literature, which is precisely the part that's hardest to self-certify in a community post.

Lean was developed by Leonardo de Moura at Microsoft Research, with the most active version being Lean 4. Behind it lies Mathlib, a large mathematical library that has formalized tens of thousands of classical theorems spanning algebra, topology, number theory, and more. This means AI-generated proofs can directly invoke already-verified lemmas rather than starting from scratch. In recent years, DeepMind's AlphaProof and numerous academic teams have adopted Lean as the "gold standard" for evaluating AI mathematical reasoning — because it transforms "mathematical correctness" from a subjective judgment into a machine-checkable binary result: pass or fail. The challenge of integrating Lean into AI workflows is that the model must output code conforming to Lean's syntax and type system, which itself is a demanding test of model capability.

How to Assess the "47-Year Record" Claim

The specific number "47 years" is highly shareable — it implies a long-standing mathematical boundary being pushed forward by AI. But with only a title and a screenshot to go on, readers should maintain a few basic critical perspectives:

• The specific problem is unknown: The post does not (in the available material) identify which mathematical problem or which record was broken, making independent verification by third parties impossible.
• "Record" is loosely defined: It could refer to a bound in combinatorial optimization, an improvement on a constant, or something stated more casually — the difference in significance is enormous.
• Reproducibility: A genuinely valuable breakthrough requires publicly shared code, prompts, and Lean scripts for others to reproduce. Without these, the conclusion remains at the level of "an interesting claim."

The rational stance is: give the Lean verification step credit, since it provides a verifiable anchor point — but reserve judgment on the grand narrative of "breaking a record" until the author provides reproducible materials.

The Real Significance of Experiments Like This

Setting aside whether the specific conclusion holds up, this post reflects a trend that is actively unfolding: individual developers, using off-the-shelf AI toolchains, are attempting to bring the kind of "compute + reasoning" workflows that once required large research institutions onto their own desktops — and claiming to do it for free. Regardless of whether the result ultimately earns academic recognition, the experimental approach itself — combining agent orchestration, automated reasoning, and formal verification — represents a plausible path for AI-assisted scientific research.

For readers tracking the AI frontier, rather than fixating on whether any single post is true or false, it's more valuable to focus on the underlying methodology: Can multi-agent collaboration genuinely improve performance on complex problems? Can formal verification become a standard mechanism for constraining AI hallucinations? These are the questions with lasting value.

Conclusion

This Reddit post is a classic case of "high buzz, low detail." It combines OpenAI Dots, agent swarms, Lean verification, and a mathematical record into a compelling package — yet the lack of verifiable specifics makes a definitive verdict impossible.

In an era of rapidly expanding AI capabilities, we should remain open to attempts like "challenging classical problems with free tools," while holding firm to the baseline requirement of "verifiable and reproducible." True breakthroughs will ultimately withstand rigorous scrutiny — both from tools like Lean and from independent peer review.