Anthropic has demonstrated that Claude can run multi‑day agent workflows without continuous supervision. The model receives a task, autonomously switches between test oracles, persistent memory and orchestration patterns, and compresses weeks of routine programming into hours. In 2,000 sessions Claude assembled a C compiler capable of building the Linux kernel; similar experiments are already underway in scientific projects.
The most striking example is a differentiable version of a cosmological Boltzmann solver written in JAX. Solvers such as CLASS and CAMB normally require months of effort to produce gradient‑enabled versions. Claude Opus 4.6 generated working code that automatically differentiates the equations for photons, baryons, neutrinos and dark matter on GPUs, without any input from a cosmology expert. This proves that even tasks outside a team’s core competence can be completed by an agent with minimal oversight.
For companies with R&D budgets in the tens of millions, such autonomous scenarios become economically viable. Migration of legacy code, rewriting numerical solvers and debugging large‑scale scientific projects can now be "launched" once and deliver a finished artifact in hours rather than years. Investment strategy shifts – instead of endlessly expanding on‑premise clusters, firms should consider hybrid models where part of the workload is handled by cloud AI agents.
Why this matters: Recalculating R&D ROI with savings on HPC infrastructure shows that subscriptions to autonomous AI agents cut development cycles from months to hours, speeding product launch and boosting competitiveness.