Anthropic Labs has announced a notable milestone at the intersection of artificial intelligence and formal mathematics. A specialized multi-agent system, built upon a research model from the Claude family, has completely formalized and computer-verified Fermat's Last Theorem within the interactive proof assistant Lean. The project represents the first computer-verified proof of Andrew Wiles' famous mathematical work. Within the formal mathematics and Lean communities, the result is recognized as a historic achievement demonstrating the viability of large-scale formal verification using artificial intelligence.
The technical scale of the undertaking significantly exceeds prior verification efforts in modern computing. Over an uninterrupted period of 11 days, the coordinated collective of AI agents generated more than 13 million lines of formal Lean code. The system handled the dense mathematical reasoning by parallelizing deeply nested theoretical structures and translating them into machine-checked formulations. Experts classify the completed work as the largest formal verification in the history of mathematics.
A decisive element in achieving the verification was the autonomous decomposition of the overall proof into thousands of individual components. Throughout the 11-day run, the multi-agent system proved over 30,000 intermediate mathematical lemmas autonomously, without requiring human researchers to manually intervene in individual deduction steps. Specialized Claude agents were tasked with closing logical gaps, validating assumptions, and linking intermediate conclusions together. Through this coordinated process, the multifaceted structure of Wiles' proof was completely grounded in elementary axioms.
This approach represents a fundamental methodological departure from conventional language model usage. Rather than relying on classical, error-prone text generation that frequently produces inaccurate conclusions in advanced reasoning tasks, the agents operated within a deterministic environment. The Lean interactive proof assistant served as an unyielding verification engine, checking every proposed deductive step for mathematical soundness. Logical errors were immediately rejected by the software, preventing flawed reasoning from propagating through the proof.
The breakthrough provides formal mathematics with concrete evidence that LLM-supported systems can successfully close ultra-complex logical deduction chains without hallucinations. While standard generative models in open-ended text generation often produce plausible but factually incorrect statements, integration with Lean binds the neural network to strict mathematical truth. The agent collective demonstrated that generative architectures, when paired with formal verification tools, can reliably complete the most demanding deductive chains.
The successful verification illustrates expanding capabilities for cooperative AI agents operating across complex scientific domains. By orchestrating multiple specialized Claude instances, the team completed a monumental formalization project that had long been considered a multi-decade challenge for human mathematicians working alone. The achievement highlights how autonomous agent systems can serve as dependable instruments for rigorous proof construction and critical software verification. It establishes a new reference point for integrating generative architectures with interactive theorem provers.