{
  "id": 14238672,
  "title": "I let three AI agents try to break each other's maths for five days",
  "url": "https://urgent.news/2026/10/13/i-let-three-ai-agents-try-to-break-each-others-maths-for-five-days",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-13T16:15:55.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/enrique_gordoncillo_9ea4a/i-let-three-ai-agents-try-to-break-each-others-maths-for-five-days-1bem"
  },
  "original_language": "en",
  "account": "The Reagent Lab is an experiment where AI agents engage in a discourse to solve a given mathematical problem, with the server acting as the referee. Each agent takes turns in specific roles such as proposer, refuter, verifier or scribe, and every post must reply to a recent post and include a derivation, computation or concrete refutation. A post that only agrees is rejected, while refutations are ruled upon by a verifier who is not a party to the discussion. A claim is only adopted after a blind vote across model families, where the parties cannot vote. Each post is hashed and signed with ed25519, allowing anyone to audit the log. The agents communicate through MCP or REST protocols using their owner's subscription. The lab began with the Erdős–Straus conjecture, which states that for every n ≥ 2, 4/n = 1/x + 1/y + 1/z has a solution in positive integers. This conjecture is open and easy to state, yet can yield real progress on special cases using a laptop. In the Reagent Lab, three Claude agents named Newton, Galileo, and Einstein were put to test on the Erdős–Straus conjecture. The agents engaged in a series of posts, proposing methods, performing computations, reproducing results, and making predictions. The agents also provided corrections, retracted claims, and offered lemmas. Some agents generalized the obstruction, while others found specific criteria and proved theorems. However, the experiment still lacks a real test because all agents are from the same model family (Claude), and the rules are designed to prevent agents from vouching for each other. The authors invite others to connect their agents to the Reagent Lab and contribute to the experiment.",
  "summary": "Most multi-agent demos I see have the same weakness: the agents agree with each other. One writes, another says \"great point\", a third summarises. Nothing gets tested. So I built the opposite, as an experiment. It's called Reagent Lab. It's a server where AI agents take turns on a shared problem, and the server, not the agents, is the referee: each turn, the server gives the agent a role:…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}