We report a methodological study of agentic AI in gravitational-wave data analysis: two systems, Claude Code (Anthropic) and Codex (OpenAI), autonomously executed the same simple end-to-end pipeline on Einstein Telescope (ET) simulated data, on shared infrastructure and without human intervention. The object of study is the behaviour, reliability and auditability of the agents, not the physics output, used here as a controlled test case. The pipeline comprises power spectral density estimation from simulated ET noise, geometric template bank generation with IMRPhenomD waveforms, matched-filter recovery of 100 binary black hole injections, results generation, and LLM-assisted production of a LaTeX manuscript in Physical Review D style. Both agents received identical specifications and resources. The experiment was run twice: first with unrealistically loud injections, then with signals rescaled to a physically motivated SNR range. In both runs the results converged, with comparable detection efficiency and template bank size. The agents, however, behaved very differently: Claude Code finished in about 3.4 minutes with silent deviations from the specification, while Codex needed about 16 minutes across explicit self-correcting restarts, including an unsolicited optimization of the matched-filter inner loop. In the second run, a subtle difference in interpreting the SNR-range instruction produced a genuine scientific divergence: Claude Code silently raised the SNR floor to 8 (100% efficiency), while Codex followed the specification literally down to SNR 7 and recorded one missed detection. We discuss the implications - speed versus auditability, silent deviation versus explicit self-correction, instruction interpretation, and intermediate data representations in multi-model pipelines - for agentic AI in scientific workflows, within the limits of a single-pipeline, two-run benchmark.

翻译：暂无翻译