PantheonOS

PantheonOS is a multi-agent framework for data science, aimed particularly at end-to-end single-cell and spatial biology analysis. It is built to reconcile a general-purpose agent framework with the demands of one scientific domain, and it runs privacy-preserving deployments on machines you control.

Its distinguishing module, Pantheon-Evolve, applies genetic-algorithm-driven agentic code evolution so agents improve the algorithms and code they run rather than only calling fixed tools. The project reports super-human performance on specialized scientific tasks from that loop.

Agents are composed into teams through several orchestration patterns: PantheonTeam, Sequential, Swarm, Mixture-of-Agents, and AgentAsTool. Messaging runs over NATS so a deployment can span machines, and Pantheon-Fleet extends that to controlling a distributed network of machines directly. You work through an interactive command-line REPL, a multi-agent chatroom UI, a hosted web application, or a desktop build. Optional components add retrieval and vector search, mobile gateway channels, and R language support, and a container image runs the system with a mounted workspace, reading factory templates from the image while project overrides persist in the mount.

A community store distributes more than 1,000 curated agents, teams, and skills for biomedical AI, installable from the UI or the command line. The project also publishes replayable trajectories: recorded multi-agent sessions you can step through in the UI message by message, file by file, and tool call by tool call. Published examples include Tangram deconvolution on binned spatial transcriptomics of E6.0 mouse embryos (51,711 spots across 6 embryos and 16,732 shared genes), design of a 1000-plex immune-oncology gene panel, mapping congenital heart disease genes onto a 3D PCW12 fetal heart from MERFISH and single-cell RNA sequencing, spatial mapping of single-nucleus ATAC sequencing with MOSCOT optimal transport, spatial ligand-receptor interaction analysis, and a comparison of cell-segmentation tools on a three-channel embryo image.