Anthropic reports that a multi-agent Claude system formalized Fermat’s Last Theorem in Lean over eleven days, with occasional high-level human direction.
This formalizes an existing mathematical proof. It is not the discovery of a new proof of an unsolved conjecture.
The work builds on an established mathematical and formalization community.
The interesting operational feature is the check at the end: a proof assistant can verify the submitted formal argument.
Many agent tasks have no comparably crisp success criterion. Sending a persuasive email and completing a correct proof leave very different kinds of evidence.
Observing capable systems means recording how success was established, as well as what the agent produced.
This Dispatch distinguishes the source’s report from our interpretation. Our editorial method.