Claude Science Completes the UV Map of the Sky: What the Agents Did

Anyone who wanted to see the sky in ultraviolet light ran into a hole: about a third of it was missing, including large parts of the Milky Way. Astrophysicist Brice Ménard has now closed that gap with help from AI agents in Claude Science, Anthropic’s research environment. The result is the first complete ultraviolet map of the sky, and the way it was made shows quite precisely what such agents can and cannot do in science today.

Key takeaways

• On October 8, 2026, Anthropic published Ménard’s account of the first complete sky map in far and near ultraviolet light.
• About two-thirds of the map is based on measurements, mainly from NASA’s GALEX mission; a model predicted the missing third from other wavelengths.
• In tests on deliberately masked regions with known values, the predictions came within about 10 percent of the real measurements on average.
• The agents handled data search, calibration and merging in a few days, yet missed a systematic image artifact in two review rounds; only Ménard caught it.
• Every pixel is labeled as measured or predicted and carries an uncertainty estimate.

Why this map did not exist until now

Ultraviolet light reveals what stays hidden in visible light, such as dust clouds lit up by young, hot stars. Earth’s ozone layer absorbs this radiation, however, so it can only be observed from space. The most important data source is NASA’s GALEX space telescope, which imaged about two-thirds of the sky in roughly 38,000 individual observations between 2003 and 2013. Its detectors were sensitive: the mission had to avoid regions with very bright stars to keep the instruments from being damaged. That is exactly where much of the galactic plane, the band of the Milky Way, lies.

Other missions have surveyed the sky in ultraviolet as well, but with different instruments, different resolutions and different quirks. Besides GALEX, Ménard lists NASA’s Swift telescope, the Korean instrument FIMS/SPEAR and the European satellite TD-1, along with data from ESA’s Planck and Gaia missions. By his account, merging these datasets into one clean picture means weeks of painstaking work. Astrophysicists understandably set such tasks aside, he writes, because more pressing research is waiting.

How the agents worked

Claude Science is the work environment Anthropic launched in beta for scientists at the end of June. A coordinating agent hands tasks to other agents, a dedicated reviewer agent checks calculations and citations, and every result comes with its code and computing environment so it can be reproduced. Ménard set the goal and steered the process. He soberly describes his own contribution as guidance; the agents did the actual implementation.

The work proceeded in several steps. First, the agents gathered publicly available ultraviolet surveys and downloaded the observations. Then they removed scattered light from bright stars so that the faint ultraviolet glow around them could be measured at all. Next, they calibrated the datasets against each other, brought them to a common resolution and mapped them onto a single coordinate system. For the missing third, Claude used a technique called inpainting: a model learns in the well-measured regions how ultraviolet brightness relates to visible light, infrared and radio emission, and infers the gaps from that. Finally, estimates of the ultraviolet light from more than 100 million individual stars were added, derived from Gaia measurements in visible light.

According to Ménard, the collaboration stretched over several days and more than a dozen successive versions of the map. Computing runs lasting hours continued without him while he turned to other projects. That is the real gain: not that a machine discovered something entirely new, but that tedious, long-postponed data work got done in a few days instead of not at all.

Where the human was needed

One episode in the account is instructive. In an intermediate version, circular patches appeared, each slightly brighter or darker than its neighbors. The circles were the footprints of individual GALEX exposures: an ultraviolet glow from Earth’s atmosphere had not been fully removed. Two review rounds by other agents had missed the error. Ménard spotted it and pointed it out, and the agents applied the correction to all 38,000 observations, which took several hours of computing time.

Just as important is what the map is and what it is not. About a third of it is a well-founded prediction, not an observation. The deviation of about 10 percent was measured on regions whose true values were known; whether the model performs equally well in the Milky Way’s plane, crowded with stars and dust, cannot be fully demonstrated that way, because that is precisely where comparison data is missing. Commentators also point out that the account so far comes almost entirely from Anthropic, where Ménard works as a researcher alongside his professorship at Johns Hopkins University; an independent review is still pending. Ménard himself claims nothing more. He sees the map mainly as a teaching tool that lets students examine the structure of the Milky Way at this wavelength in detail, and the pixel labels show every user where measurement ends and modeling begins.

What researchers can do with this

For scientists, the example is more interesting than the map itself. It points to a type of task particularly well suited to agents: many heterogeneous, publicly available datasets, clear physical criteria for what a good result looks like, and a human who knows the field well enough to spot errors at a glance. Similar cases exist in almost every discipline, from reconciling old climate records to merging genome databases. How other labs are experimenting with agent-driven research environments was recently shown by NVIDIA with its AI-built simulation worlds.

If you want to try Claude Science, you need a paid Claude plan: the beta is open to Pro, Max, Team and Enterprise customers, and in organization accounts an administrator has to enable it. The application runs on macOS and Linux, either locally or over SSH on a remote machine, all the way to the login node of a high-performance computing cluster. Its more than 60 preconfigured skills and data connectors are aimed mainly at the life sciences and chemistry so far; Ménard’s project shows that the tool can be used beyond those fields as well.

Bottom line: a tool for work left undone

The ultraviolet map is not a breakthrough in fundamental research, and it does not pretend to be. It is something more down to earth and perhaps more consequential: proof that data work nobody wanted to do for years can be finished by agents in days, as long as an expert checks the results with a trained eye. Unlike the contested AI proofs in mathematics, the question here is not whether a machine produces new insights, but the many preparatory steps without which insight never gets started. If Ménard is right, projects like this are sitting in drawers at many institutes.