AI and the World's Hard Problems · Astronomy & space · Anthropic · 2026-10-09
Long-running Claude for scientific computing
An Anthropic researcher explains how to apply multi-day autonomous agentic coding workflows, using test oracles, persistent memory and orchestration patterns, to scientific computing. As an example, Claude Opus 4.6 autonomously built a differentiable version of a cosmological Boltzmann solver, which predicts the statistical properties of the Cosmic Microwave Background, over several days, compressing months of work. The approach suits well-scoped tasks with clear success criteria, such as reimplementing numerical solvers or converting legacy Fortran code.