AI Agent Solves Scientific Computing Tasks on Researcher Laptops Without Facility Infrastructure

Scientists typically spend considerable time translating their research goals into concrete computing tasks—configuring environments, requesting resources, submitting jobs, and then manually combing through logs to identify what went wrong. A new system called ASCEND (Autonomous Scientific Computing Engine and Novel Discovery) aims to automate this friction-filled workflow by putting an AI agent at researchers' fingertips.

How ASCEND Works#

Rather than requiring a centralized computational facility to host intelligence, ASCEND runs an AI agent on a researcher's own laptop. The agent communicates with remote Slurm-managed high-performance clusters and GPU workstations through an authenticated, multiplexed connection. The language model—either Claude Code or Codex—is hosted remotely and proposes actions, but it never holds any credentials. A locally executed policy checker ensures site rules are respected before any request reaches a cluster. Individual user accounts on each resource are all that's needed; the system provides a public installer so researchers can connect their own infrastructure.

Real-World Case Studies#

The team documented four scenarios where ASCEND successfully handled complex tasks:

• Error recovery: The agent diagnosed a simulated hardware fault on a GPU, repaired the code, and resubmitted the job while preserving intermediate artifacts.

• Reproducing published research: Working from released forecasts, ASCEND recovered an undocumented evaluation protocol and reproduced published weather-forecasting results to within 2.1 percent on one metric and 2.4 percent on another.

• Parallelization under constraints: The system accelerated a 12,693-line geophysics solver from roughly twelve hours to a fraction of that runtime while maintaining bit-for-bit numerical identity with the original serial version.

• Bug discovery: During parallelization, the process uncovered two instances of undefined behavior in the original code; both were fixed and reported to upstream developers.

Safety and Reliability#

The deployed policy validator correctly rejected 29 of 30 artificially constructed violations and flagged one for human review. However, it also denied three of fourteen legitimate requests, suggesting the current safety guardrails are still being refined. The work emphasizes that autonomy was exercised under human supervision during testing; a fully autonomous recovery benchmark remains a future goal.