The narrative around Large Language Models often devolves into debates about code quality or existential dread, but a recent case study from Spain suggests a more pragmatic application: astronomical discovery. Pavel Rabtsevich, a 28-year-old product manager with no formal astronomy background, utilized Anthropic’s Claude Code alongside OpenAI’s Codex to analyze data from NASA’s Transiting Exoplanet Survey Satellite (TESS). By processing measurements from over 126,000 stars, Rabtsevich identified a candidate exoplanet around star TIC 4206066, a find that has since been accepted for official observation by the scientific community.

The Agentic Workflow

Rabtsevich’s methodology was not a simple prompt-and-answer session. He fed public TESS data, captured between November 2024 and January 2025, into TypeSafe’s Jev, Claude Code, and Codex. He explicitly rejected the models' initial outputs, engaging in iterative refinement rounds to improve detection methods. "I didn’t take the models’ first answers," Rabtsevich told Gizmodo. "I pushed them, round after round, to improve their own methods, and checked what they gave me against other models." This human-in-the-loop approach was critical; Rabtsevich retained judgment on which tests to run and which results were statistically significant enough to keep.

Validating the Signal

The candidate, TIC 4206066, is a star slightly smaller than our Sun, located approximately 116 light-years from Earth. The AI-assisted analysis revealed a consistent dip in stellar brightness every 3.18 Earth days. To rule out transient anomalies, Rabtsevich directed the models to cross-reference this pattern against historical TESS data from 2018 and 2020. The periodic dimming appeared in all datasets, strengthening the hypothesis of a planet with a radius just over 1.4 times that of Earth. He ran seventy-four distinct tests with Claude Code to attempt to falsify the hypothesis, including a predictive test where the model correctly forecasted dimming times across three separate datasets.

Democratizing Discovery

While professional astronomers have cataloged over 6,400 exoplanets, the volume of available data is exploding. Kevin Hardegree-Ullman, a research scientist at the NASA Exoplanet Science Institute, noted that AI tools are becoming essential for vetting candidates, though confirmation remains a human and hardware bottleneck. "Without these tools I couldn’t have processed a data set this size and carried it through to a result," Rabtsevich said, contrasting his two-week AI-assisted workflow with the months of spreadsheet grinding required just a few years ago. The tool’s ability to string together publicly available code and run basic vetting processes allows amateurs to contribute meaningfully to the field.

Key Takeaways

  • Iterative Refinement is Key: Rabtsevich did not accept first-pass AI outputs, using multi-round prompting to improve detection logic.
  • Cross-Validation Matters: The discovery was strengthened by checking historical data (2018/2020) against recent findings (2024/2025).
  • Human Oversight Remains Critical: The amateur astronomer retained final judgment on statistical significance and test selection.
  • Official Recognition: MIT approved the request for TESS to observe the star from October 31 to November 26, 2026.

The Bottom Line

This isn't about AI replacing scientists; it's about AI removing the computational friction that previously excluded non-specialists from high-level data analysis. If TESS confirms the transit, we have a new benchmark for how agentic coding tools can accelerate scientific discovery outside of traditional institutional walls.