A Product Manager, Three AI Agents, and a Possible Planet 116 Light-Years Away

Pavel Rabtsevich is 28 years old, works as a product manager in Spain, and has never studied astronomy. He does not own a telescope. What he had instead was an internet connection, subscriptions to several AI tools, and a hunch about NASA data.

Last month he handed a public dataset from NASA's Transiting Exoplanet Survey Satellite, TESS, to three AI coding agents: TypeSafe's Jev, Anthropic's Claude Code, and OpenAI's Codex. The data covered a narrow slice of TESS observations, more than 126,000 stars measured from November 2025 into early January 2026. He asked the agents to hunt for the telltale flicker of exoplanets, a star dimming at regular intervals as a planet crosses its face. Round after round, he pushed the models to improve their methods and cross-checked each model against the others.

One star survived the gauntlet. TIC 4206066, a star slightly smaller than our Sun located about 116 light-years away in the constellation Orion, dimmed for roughly two hours every 3.18 days. The pattern was clean enough that Rabtsevich kept digging. In TESS archives from 2018 and 2020 he found 22 more instances of the same dimming, at the same cadence.

The planet that might be there

If the signal is planetary, the numbers describe a harsh world. Rabtsevich calculates it would be a super-Earth roughly 1.4 times the radius of our planet, circling its star once every three days. At that distance the estimated surface temperature reaches about 1,000 degrees Fahrenheit, or 540 degrees Celsius. That is hotter than Mercury, the innermost planet of our solar system, and hot enough to melt lead. His analysis also picked up a second, fainter dimming pattern with an 11.13-day period, possibly a second planet, but he considers that evidence far less convincing.

The AI did the plumbing, the human did the science

Rabtsevich is careful about who gets credit. The models wrote the code, ran it, and checked each other's work, he told Scientific American. But the research itself was his: the question, the plan, the rules, the computing, the decisions on every candidate, and what to publish.

He also stress-tested the result like an adversary. In total he ran 74 tests with Claude Code aimed at disproving the hypothesis. In one experiment he gave the agent data from two observing runs and asked it to predict when the dimming would appear in a third. The predictions held in all three cases.

The professionals had seen it and moved on

Here is the part that makes astronomers pay attention. Pablo Mercader Perez, a PhD student at MIT on the TESS team, says a detection filter flagged this exact signal in 2020 and rejected it for lack of conclusive evidence. Since then TESS has observed that sector of sky many more times. With the added data, the signal grew clearer. In terms of statistical significance, Mercader Perez says, the candidate became effectively detectable only in January of this year.

Rabtsevich was not even the only machine to notice. A deep-learning model built by the team of Mercader Perez independently flagged the potential planet. So did LEO-Vetter, software developed by astronomer Michelle Kunimoto, which selected the candidate back in early 2026. Greg Laughlin, an astronomer at Yale who now works in the AI industry, calls the evidence for at least one planet around the star solid, and the work a great example of human-AI teamwork.

Based on the analysis, MIT approved a request from Rabtsevich for dedicated follow-up. From October 31 to November 26, TESS will measure the brightness of the star every two minutes, a high-cadence stare designed to settle the question. Rabtsevich has published the exact times the star should dim during that window on Zenodo. As he put it in a post on X, if the star dims on time in November, this gets a lot more serious. If it does not, he will post that too.

The honest caveats

Stars dim for many reasons that have nothing to do with planets. Rabtsevich noticed that the 2018 dips tended to align with moments when TESS fired its thrusters, which may have skewed the measurements. Binary star systems can mimic planetary transits. All that is confirmed today, he says, is a repeating transit-like signal in open data, with a plausible planetary explanation and alternatives that are not yet ruled out.

Kevin Hardegree-Ullman, a research scientist at the NASA Exoplanet Science Institute, describes the core problem: confirming planets means ruling out false positives, and that takes scarce telescope time that astronomers compete for. AI, he notes, may be less useful for the confirmation step than for the initial search.

No telescope, no degree, no lab. Just an internet connection, three AI agents, and 126,000 stars. Whether or not the planet is real, the pipeline is.

Why this matters beyond one star

The professional community has already quietly handed much of the searching to machines. An AI model NASA deployed roughly a year ago to comb TESS data has already identified around 7,000 exoplanet candidates. Astronomers have cataloged more than 6,400 confirmed exoplanets in total, and TESS alone has produced over 1,000 confirmed candidates in its eight years of scanning. The bottleneck is no longer collecting the data. It is deciding which blips deserve a closer look.

That is exactly the step Rabtsevich automated, from his desk, with off-the-shelf tools. Since he published his findings, other amateurs have begun their own AI-assisted dives into TESS and other NASA archives, and some report promising early results. The worry, which Rabtsevich himself voiced, is that enthusiasm without careful checks could flood professionals with poorly vetted signals. But the deeper point stands: the tools that once required a PhD and a supercomputer now fit into a hobbyist workflow. The next census of worlds may be written by people who never owned a telescope.