A lot of the work in spectroscopy is not hard so much as slow. Someone has to look at each line fit, decide whether it is good, and fix it when it is not, and on a survey of a few hundred thousand stars nobody can. Language models are now good enough to make some of those calls. I build agents and tools that do this inside real analyses, and I try to be strict about which decisions they should and should not be making.
Egent (Ting, Saad, Liu & Shen, 2025) measures equivalent widths of stellar absorption lines, one of the most labor-intensive steps in detailed abundance work. It fits each line with multi-Voigt profiles, then a language model looks at the fit, catches the failures and refits them, working directly on the raw flux spectra. Compared with expert measurements for 84 stars and more than 18,000 lines, it matches the human values and takes days instead of months. It runs from a web interface or offline on a laptop.
In Saad & Ting (2026), instead of writing one fixed pipeline, I split a spectroscopic analysis into small tool servers that an agent can call and combine, with the operating decisions written down in a single instruction file. The first application is a search for double-lined binaries in APOGEE DR19, which returned 41,466 candidates among 238,205 dwarfs (the science is on the Binary Stars page). The agent drives the benchmark and the worked examples, and the full catalog run is the same deterministic core under a driver script, so the result does not depend on the model behaving well 238,000 times. Code on GitHub.
AGN-Egent applies the same idea to quasars. It separates broad and narrow emission lines in SDSS and DESI spectra, estimates black hole masses, sends the fits that fail its quality checks to an inspector that picks one fix from a fixed menu, and ranks a whole survey by how unusual each object is. There is a browser demo.
Many decisions during an observing night have a short list of possible answers: is the focus good, is the sky clear enough, which target next. SkyJev (Saad et al., in preparation) builds that list from the measurements and has a small open model assign a probability to each option in a single pass, with no text generation. It acts only when it is confident and passes the rest to a larger model with tools, or to the observer. Each choice takes about 0.1 s. On 300 archival focus sequences it had not seen, it acted on 69% of the focus decisions and agreed with a fit to the full sequence in 97% of those. I am testing it on decisions from real nights at the Large Binocular Telescope.
astrodata-mcp is an MCP server that lets any compatible AI client query public archives directly: the ESO Science Archive (including HARPS, ESPRESSO and UVES), the Keck Observatory Archive, Gaia DR3, SIMBAD, VizieR and published Magellan catalogs. All tools are read-only and need no credentials.
The same models are also changing how astronomy papers are written. With Yuan-Sen Ting, I counted the vocabulary that language models favor in the full text of about 220,000 astro-ph papers from 2015 to mid-2026, calibrated it on papers whose authors declared model use, and inferred the assisted fraction each year with a hierarchical Bayesian model. Our estimate is that more than half of recent papers were written with some model assistance. The model is built for the population, not for judging any single paper, so we do not release per-paper scores. Code and derived data are on GitHub.