VISION-CC (Virtual Scientific Companion with Claude Code) uses a team of Claude Code agents to run autonomous experiments (AE). The scientist steers the experiment with evolving goals; a team-lead agent relays their direction to four cogs (specialized agents): op-cog drives the beamline, ana-cog analyses the data, gpcam-cog decides where to measure next with gpCAM Bayesian optimization, and writer-cog provides the visualization. An audit-cog reconstructs offline, who proposed, decided, and corrected what during a campaign.
VISION-CC succeeds VISION, our earlier LLM-based assistant, moving from a single LLM to a team of Claude Code agents.
| Path | What |
|---|---|
CLAUDE.md, .claude/, docs/ |
the agent team: team-lead procedure, cog definitions, skills, and shared protocols |
src/audit/codebook.md |
the audit's definitions and the exact prompt of every model step |
data/exp2_p07_s44/ |
44 SAXS frames from Exp. 2, losslessly compressed, with index, mask and decoder (frames.py) |
results/ledgers/exp2/ |
an example audit ledger: 25 minutes of Exp. 2 in which the scientist proposed and steered the peak-07 mapping that produced these frames (sanitized) |
Read a frame: frames.load("data/exp2_p07_s44/" + row["frame"], (1679, 1475)) for a row of index.csv.
More code will follow in this repository.
Paper: G. Luthra, R. J. Pandolfi, B. T. Diroll, R. Ronquillo, R. Li, C. B. Murray, M. M. Noack, K. G. Yager, E. H. R. Tsai, Auditable Human-Agent Teams for Autonomous X-ray Experimentation, arXiv (2026, link to follow). Contact: Esther Tsai (etsai@bnl.gov)