Long-running Claude for scientific computing

In an Anthropic Research post, Siddharth Mishra-Sharma outlines how to apply multi-day agentic coding workflows to scientific computing tasks using Claude Code. Traditionally, scientists use AI agents in tight conversational loops, but improvements in long-horizon tasks allow users to specify high-level objectives and let autonomous agent teams work for hours or days. As a concrete example, Mishra-Sharma used Claude Opus 4.6 to implement a differentiable version of a cosmological Boltzmann solver in JAX, a task outside his core scientific domain. Because a Boltzmann solver is a deeply coupled pipeline requiring sequential debugging and domain knowledge, the project relied on a specific setup using an HPC cluster with the SLURM scheduler. The workflow relies heavily on a CLAUDE.md file in the root directory where instructions and project deliverables are articulated and updated by the agent. It also uses a CHANGELOG.md file as a portable long-term memory to track progress, completed tasks, failed approaches, and accuracy tables. To ensure the agent stays on track, the process uses test oracles such as the CLASS C source code reference implementation, continuous unit testing, and Git commits after every meaningful unit of work. Users can orchestrate this by launching Claude Code inside a tmux session on a compute node, allowing them to detach, monitor progress via GitHub, and intervene only when necessary.