Anthropic4 Sept 2026Anthropic: Formalizing Fermat’s Last TheoremFermat’s Last Theorem now has a complete computer-checked proofClaude spent 11 days translating a monumental mathematical result into something Lean could verify, drawing on decades of human work.
Anthropic1 Sept 2026AnthropicSecurity monitoring with the records under customer controlAnthropic’s proposed safeguards separate who stores sensitive activity from who detects and reviews misuse.
Anthropic27 Aug 2026AnthropicA shared interface for laboratory equipmentAnthropic and HHMI Janelia are testing a standard that lets agents inspect device state and operate scientific instruments.
Anthropic10 Aug 2026Anthropic: Claude’s progress on the Riemann hypothesisA substantial step on the Riemann hypothesis, with a precise limitAn unreleased Claude model raised a longstanding mathematical lower bound from 41.6% to 67.2%. The distinction between that result and solving the famous conjecture is essential.
Anthropic25 Mar 2026AnthropicWhen users approve 93% of prompts, permission design has a signal problemClaude Code’s auto mode tried to reserve scrutiny for actions that could exceed the user’s intent.
Anthropic21 Jan 2026AnthropicAnthropic kept redesigning a hiring test because its own models could pass itA performance-engineering exercise reveals the changing meaning of a strong take-home submission.
Anthropic26 Nov 2025AnthropicLong-running agents need a handover that survives their own memoryA 2025 engineering account treats each context window like a new shift.
Anthropic4 Nov 2025AnthropicA tool result does not always need to pass through the modelExecuting code around MCP changes where data gets processed.
Anthropic16 Oct 2025AnthropicAgent skills turn specialist procedure into something that can be loaded on demandThe original Skills design is an onboarding mechanism for software workers.
Anthropic27 Jun 2025AnthropicThe AI shopkeeper could find tungsten cubes but could not reliably protect its marginProject Vend exposed a conflict between being helpful and running a business.
Anthropic26 Jun 2025AnthropicInstalling a local AI tool should not require becoming its developerDesktop Extensions addressed the last mile between an MCP demo and ordinary use.
Anthropic8 Apr 2025AnthropicA student’s chat history cannot by itself tell you whether learning happenedAnthropic’s first education report separates observed requests from educational outcomes.
Anthropic20 Mar 2025AnthropicA deliberately uneventful tool gave an agent room to reconsiderThe “think” tool is also a lesson in how quickly AI engineering advice ages.
Anthropic19 Dec 2024AnthropicAn agent architecture begins with deciding who controls the next stepAnthropic’s 2024 guide distinguishes fixed workflows from systems that choose their own path.
