AI research agents and the ultraviolet sky map: what the expert caught

On October 8, Anthropic described a project in which Johns Hopkins astrophysicist Brice Ménard used Claude Science, its research platform, to assemble what it calls the first complete map of the sky in ultraviolet light. About a third of that map was never observed by a telescope. It was inferred. This matters well beyond astronomy, because it is a clear case of research agents doing long, multi-step work, and of how people can tell which parts of the result to trust.

What was built

Most ultraviolet data comes from NASA's GALEX mission, which took about 38,000 observations between 2003 and 2013 and covered roughly two thirds of the sky. It avoided very bright regions, including much of the plane of the Milky Way, to protect its detectors. Those gaps stayed open for years.

According to the reports, AI agents handled the heavy lifting: downloading archive data, removing artifacts, calibrating different instruments to a common scale and aligning coordinates. The missing third was filled by a statistical model that predicts ultraviolet brightness from measurements of the same regions at other wavelengths, such as visible, infrared and radio. Ménard guided the work over several days.

Two habits worth copying

The first habit is honest labelling. Every pixel in the map is marked as observed or predicted, and each carries an uncertainty estimate. A reader can see at once where the data ends and the model begins.

A complete map is not the same as an observed one. The useful thing is that this map tells you which is which.

The second habit is testing against known answers. Ménard hid regions that already had real measurements, asked the pipeline to rebuild them, and compared the results with the truth. After several rounds of refinement, the reports say predictions came within about 10 percent of the hidden measurements.

That test has a limit, and the reports are open about it. The masked regions were areas the telescope had observed. The real gaps sit around bright stars and along the plane of the Milky Way, where the sky looks different. A model that does well on the test regions may do worse where it matters most. One report notes that the documentation warns errors grow with distance from observed data, and that the validation had not yet been peer reviewed.

What the review agents missed

The most telling detail is small. Intermediate versions of the map showed faint circles in the darkest fields. Each circle was the imprint of a single old exposure, where glow from Earth's upper atmosphere had not been fully removed. The agent had flagged this glow as a known risk early on. Yet, according to the reports, two rounds of review by agents did not catch its traces in the combined map.

Ménard did. He spotted the circles by looking at the map and asked for a fix. The system then reprocessed all of the roughly 38,000 observations in a few hours. The speed of the fix is impressive. The way the problem was found is the lesson.

An expert who knows what the sky should look like noticed a pattern that did not belong. Automated reviewers, checking the same work, did not. This is not a story about agents failing. It is a story about where human judgment still adds the most: at the point where something looks plausible but is subtly wrong.

What this means for deep research

Research agents in business and science face the same questions. Which claims come from sources, and which are the agent's inference? How was the method tested, and does the test resemble the real cases? Who with real expertise looked at the output before anyone relied on it? A few practical checks follow:

• Mark every finding as sourced or inferred, and keep that label wherever the result travels.
• Test the method on cases where the answer is already known, then ask how those cases differ from the ones you care about.
• Have someone who knows the field inspect the result directly, not only a summary of it.
• Turn each error an expert catches into a new check that runs on the next project.

In practice

This project shows research agents doing serious work at a speed few teams could match, with a scientist steering. It also shows that capability and reliability are separate questions. The map is trustworthy because it is labelled, tested against real measurements and inspected by an expert who caught what automated review missed. At SoReliable, we think that pattern should be the standard for any team that relies on AI research: measured expert judgment, built into the workflow, not added at the end.