Pavel Rabtsevich has no astronomy degree and no telescope. He is a 28-year-old product manager in Tarragona, Spain. What he had was an internet connection, several AI subscriptions and free access to data from TESS, the NASA space telescope that has been watching the sky for planets since 2018. In September he used AI coding agents to dig through that data, and found a signal that may be a planet nobody had noticed.

How he did it

TESS finds planets by measuring the brightness of stars. When a planet passes in front of its star, the star dims slightly, and if the dip repeats on a regular schedule, it may be a planet orbiting it. Rabtsevich took the data from one TESS observation period, which ran from November last year to early January and covered, by his count, more than 126,000 stars.

He handed the same data to three AI tools: Anthropic's Claude Code, OpenAI's Codex and Jev from TypeSafe. These are not chatbots but agents: they write and run their own code, download data and repeat the analysis until it works. Rabtsevich asked them to look for regular dips in brightness, narrow down the list of candidates and improve their own methods, then compared what the different models found.

The work split clearly. The AI wrote and ran the analysis. Rabtsevich decided which checks to run, whether a result was convincing and what to keep.

What he found

The strongest candidate is around the star TIC 4206066, a K-type dwarf slightly smaller than the Sun, about 116 light-years away. Every 3.18 days the star dims, which fits a planet about 1.4 times the size of Earth, with a year lasting just over three days and a surface temperature estimated at around 1,000°F (about 540°C). A second, weaker signal repeats every 11.13 days and could be a second planet, but Rabtsevich himself calls it more tentative.

Then he tried to prove himself wrong. With Claude Code he ran 74 tests designed to break the hypothesis. In one of them, the agent used two sets of observations of the star, including archive data from 2018 and 2020, to predict when it should dim in a third. The predictions held in all three cases.

A test written before the answer exists

The most scientifically careful step came next. On October 6, Rabtsevich published a protocol on Zenodo, the open research archive, laying out exactly what should happen when TESS observes the star again, in observation period 110, from October 31 to November 26, measuring its brightness every two minutes. The document gives nine predicted dimmings for the main signal and three for the second, each with a margin of about four minutes.

Because the predictions were fixed before the new data exists, nobody can fit the answer afterwards. The protocol also sets four possible outcomes in advance: signal recovered, inconsistent, excluded, or not enough data. Rabtsevich is careful about what a success would mean: even if the star dims on schedule, that alone would not prove it is a planet. According to Gizmodo, his request to MIT, which runs TESS operations, to measure the star at that two-minute rate was approved.

What scientists say

Kevin Hardegree-Ullman, a research scientist at NASA's Exoplanet Science Institute, told Gizmodo that finding a signal is the easier part. The hard part is ruling out false alarms, such as two stars orbiting each other, which can produce similar dips. In his view, AI tools won't necessarily help with that confirmation step.

Rabtsevich agrees that a match does not prove the cause. He told Gizmodo that if the star dims on time in November, "this gets a lot more serious." More than 6,400 exoplanets have been catalogued so far, and TESS candidates have led to more than a thousand confirmed planets, but each one needed professional astronomers and extra observations.

What this means if you use AI for your own data

Choose an agent, not a chat, for real data work. A chatbot can explain a method. An agent such as Claude Code or OpenAI's Codex can download thousands of files, write the analysis, run it, see the error and fix it. That is what let one person process data on 126,000 stars.

Check one model with another. Rabtsevich gave the same data to three tools and compared their answers. When two independent models arrive at the same result, it is a stronger signal than one model being confident.

Ask the AI to prove you wrong. The 74 tests were designed to break the idea, not to confirm it. The same approach works for a business forecast or a market analysis.

Write down your prediction before you look. Publishing the predictions in advance is the most valuable habit in this story, and it works with any AI: it makes it impossible to quietly adjust the answer to fit.

We will report what TESS shows after November 26.