Claude Agents Map the Full UV Sky, Predicting a Third of It

Anthropic published a research post on 8 October 2026 in which Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, describes using Claude Science to produce what the post calls the first complete map of the sky in ultraviolet light.

About a third of the map, including much of the galactic plane, was predicted rather than observed, the post says. Additional layers of the map label each pixel as “measured” or “predicted” and provide uncertainty estimates.

Why the UV Sky Had Holes

UV light is absorbed by the ozone layer, so it has to be observed from space, the post explains. Ménard writes that when he teaches astrophysics he has always had to apologize, because “The only UV map I could show was full of holes.”

The largest dataset, the post says, comes from NASA’s GALEX mission, which ran from 2003 to 2013 and imaged about two-thirds of the sky in some 38,000 separate observations. GALEX deliberately skipped locations with very bright stars, including the plane of the Milky Way, because of the risk of damaging its detectors.

Other telescopes, such as NASA’s Swift and South Korea’s FIMS/SPEAR, added data, but the post says even the combined datasets have missing pieces.

The Work Nobody Had Time For

Statistical techniques to estimate the missing data already exist. The post says doing this properly “takes weeks of painstaking work, involving meticulous calibrations at the level of individual pixels and sophisticated, repeated analyses.”

Ménard describes this as backlog work: projects that would explain a key concept or help other researchers but never rise high enough on anyone’s list. He writes: “With Claude, it has become easier to tackle such lower-priority work.”

Claude Science is the workbench Anthropic released in beta on 30 June 2026 for Claude Pro, Max, Team and Enterprise users, according to Anthropic’s launch post. That post describes a generalist coordinating agent that can spin up other agents, and a reviewer agent that checks citations and calculations. The UV map post appeared on the same day the White House announced AI-for-science compute pledges, which we covered separately.

What the Agents Did

Ménard says he started this summer with instructions that were “simple to state, but not to carry out”: gather every available UV dataset, put them on a common scale, merge them into one map and fill every patch of sky no UV telescope has observed.

According to the post, Claude orchestrated a team of AI agents. They searched the web for public UV surveys, downloaded the data and made each survey internally consistent, removing the glare that bright stars add to nearby observations. Many agents worked in parallel on different regions of the sky.

The surveys were then cross-calibrated against one another, redrawn at the same resolution and mapped onto a common coordinate system. Ménard says he gave high-level instructions and Claude set a team of agents on the tasks.

How the Missing Third Was Filled

Roughly a third of the sky has never been observed in UV, the post says. Ménard asked Claude Science to use inpainting, a machine-learning technique that restores missing parts of an image from their surroundings.

Claude combined that with observations the team did have at other wavelengths, including visible, infrared and radio. Using the two-thirds of the sky mapped in UV, it learned how UV brightness relates to those wavelengths, then applied the relationship to the unobserved third and estimated how confident it was at each point.

To test the method, Claude hid parts of regions that already had UV data and predicted them without the real values. The post says that after several rounds of refinement the model estimated the hidden data “to within about 10% of the real UV measurements”. On top of that background, Claude added estimates of UV light from more than 100 million stars inferred from ESA’s Gaia satellite.

Where the Agents Missed

The post is specific about an error. Looking through processed images one evening, Ménard noticed faint circles in one of the dimmest fields: the footprints of individual GALEX observations, each carrying leftover glow from Earth’s atmosphere.

Claude had listed this as a known issue at the start, the post says, but the map had still passed two rounds of review by other agents without the problem being caught. After Ménard pointed it out, the agents traced the cause and Claude corrected the glow in all 38,000 observations; the circles were gone after a couple of hours of processing.

Who Did What

Ménard says he and Claude produced more than a dozen successive versions of the map over several days. He would plan the next steps in a few exchanges, then Claude would run hours of computations on its own while he worked on other projects.

The measurements come from GALEX and Swift (NASA), FIMS/SPEAR (Korea), TD-1 (Europe), and Planck and Gaia (ESA), the post’s acknowledgments say. Ménard writes: “My contribution was limited to guiding them in the process.”

The Structural Read

The post frames the map as backlog work: a resource that takes weeks of calibration and never rises high enough on a researcher’s list. The change it describes is in who does that labour, not in the statistics, since the post says the techniques already existed.

Verification ran two ways. The post says the model was tested by hiding regions that had real UV data, landing within about 10% of the measurements, and the finished map labels every predicted pixel as predicted with uncertainty estimates.

The miss is the instructive part. Claude had listed the atmospheric glow as a known issue, the post says, yet the map passed two rounds of agent review before the researcher saw faint discs in a dim field and asked for a fix.

Three Implications

RESEARCH LEADS The post’s division of labour is explicit: high-level instructions and review from the researcher, hours of computation run by the agents in between.

TEAMS DEPLOYING AGENTS Two rounds of agent review passed a known flaw that a human then spotted, according to the post, so human checkpoints stayed in the loop.

READERS OF AI-FOR-SCIENCE CLAIMS The accuracy figure is a test on regions with real data, and the predicted third is labelled as predicted on the map itself.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Product Overhang Doctrine.

The framework opens: “In any product cycle, there are two curves to follow. The first is what your underlying technology can do. The second is what your users can actually do with it.”

As we read it, the UV map is a case of a user closing that gap on one project: the post describes a workbench turning a weeks-long calibration job into days of guided agent work, with the researcher still checking the result.

What Is Not Established

The 10% figure is the post’s own test on regions that already had UV data; the predicted third is an estimate, and the map marks those pixels as predicted with uncertainty estimates. The post is a first-person account published by Anthropic, which makes Claude Science.

We could not open the interactive map: the Johns Hopkins page returned an access check to our reader on 10 October. Nothing here says how the method would perform on other surveys or other fields.

The Bottom Line

Anthropic’s 8 October post says Claude Science agents gathered, calibrated and merged public UV surveys and predicted the roughly one-third of the sky never observed in UV, testing to within about 10% on hidden data. A human caught an error that two rounds of agent review missed, and the map labels every predicted pixel as predicted.