AI Agents Solved Thomson Geometry Problem
Ten Claude Sonnet 5.5 agents produced a verified formal proof for a century-old configuration challenge.
Updated on Sept. 30, 2026 in Mathematics
Live Poll
Would you trust an AI-generated mathematical proof that has not been reviewed by a human expert?
In September 2026, a swarm of ten Claude Sonnet 5.5 agents completed a 17,895-line formal proof for the Thomson problem. The result confirms the pentagonal bipyramid as the lowest-energy configuration for seven points on a sphere.
Why it matters
This experiment demonstrates the ability of multi-agent systems to coordinate on complex, multi-step logical tasks by partitioning mathematical research. It marks a shift from human-only formalization to automated agents handling intensive deductive workloads.
The agents generated 47,854 declarations over 15 hours, exchanging 1,270 messages to coordinate. The resulting proof was validated by both the standard Lean kernel and an independent implementation called nanoda.
The players
Vals AI
An organization focused on demonstrating the capability of large language models on complex coding and mathematical tasks.
Claude Sonnet 5.5
An AI model architecture capable of multi-agent coordination and complex logic generation.
The details
The agents solved the problem by partitioning the configuration space based on the smallest inner product between charge pairs. They employed semidefinite bounds—a mathematical optimization method using matrices—and three-point certificates to rigorously cover all geometric cases. By utilizing established techniques from prior human mathematical research, the agents constructed a formal proof in Lean, a programming language and proof assistant designed for verifying mathematical theorems.
Timeline
1904: J.J. Thomson originally proposed the geometry problem.
September 2026: The agent swarm performed the calculations over 15 hours.
The Tech Race
This effort follows the trend within the Lean mathematical formalization project to automate the verification of classic geometry challenges. It marks a departure from human-led formalization by demonstrating that coordinated AI agents can handle large-scale deductive proofs.
This development serves as a proof-of-concept for researchers looking to automate the verification of complex logical workloads. While the current proof remains unreviewed, the methodology offers a new template for professional mathematicians and software engineers to use AI as a formal verification assistant.
The takeaway
This work establishes that multi-agent LLM systems can successfully navigate the rigorous requirements of formal logic assistants. Observers should track whether similar autonomous agent approaches are successfully applied to unsolved mathematical conjectures or verified against peer-reviewed standards.
Further reading
Explore the latest developments in automated logical verification at Mathematics.
More information
Review the technical files and proof code at the GitHub repository for proof.
Source note: This article includes information reported by Startup Fortune.
Live Poll
Would you trust an AI-generated mathematical proof that has not been reviewed by a human expert?
