Long-Running Claude for Scientific Computing

In scientific computing, it can be effective to treat Claude Code not as a conversational assistant but as an autonomous agent that works on its own for days at a time. For problems with a clear scope and well-defined success criteria, humans don't need to keep giving detailed instructions — a high-level goal is enough.

In this approach, the key before starting is to write the project goals, design principles, and accuracy criteria into CLAUDE.md. While work is in progress, a CHANGELOG.md serves as "portable long-term memory," recording progress, failed approaches, and remaining tasks.

To judge whether the agent is actually making progress, you need a test oracle. As in the example in this piece, for scientific code this role is played by reference implementations, clear numerical targets, and test suites — and when needed, the agent is instructed to add more tests and keep running them.

Git can also be used as a coordination tool. By committing whenever a meaningful unit of work is finished, and only reflecting tests that pass without failures, progress remains visible even during long-running work, and recovery is easy if things are interrupted midway.

The execution environment was an HPC cluster using SLURM, but the core pattern is similar anywhere. Claude Code is run inside tmux, reconnecting when needed to adjust instructions, and orchestration such as a Ralph loop — which checks whether the work is "really done" — is used to prevent premature termination.

As a concrete example, the piece introduces a task implementing a differentiable cosmological Boltzmann solver in JAX. This code describes the CMB, the afterglow of the Big Bang, and targeted an accuracy of about 0.1% against existing reference implementations like CLASS and CAMB.

Importantly, this task was outside the author's core area of expertise. Even so, Claude kept working for several days based on the references, tests, and progress log, reaching sub-percent agreement across multiple outputs.

The results aren't yet production-grade in every area, but they show that for well-defined scientific computing problems, agentic development can compress human research time from months to days.