Pando: A Living Research Environment for Long-horizon Agentic Science

Recent advances in artificial intelligence (AI) are fundamentally reshaping how scientific research is conducted. AI agents increasingly plan, execute, and interpret experiments spanning many sessions, yet these long-horizon runs depend upon two properties that remain improvised rather than engineered: persistent memory that survives context resets and reproducibility that does not depend on discipline. We present Pando, a living research environment that makes both properties structural rather than aspirational by mechanically enforcing organization through the pando command-line interface (CLI). A Pando project is a git repository with a fixed shape, operated only through that CLI, which manages the project's memory, commits on the operator's behalf, and refuses the operations that would corrupt the record. We stress-tested the environment with a 24-hour autonomous de novo protein binder design campaign. The agent performed 1,404 runs across 29 experiments, authored all 4,384 commits, and delivered the designed sequences the prompt demanded. Every completed run carried a logged result, a provenance stamp identifying what executed, and a hash manifest of the outputs. This resulting record supported audits and corrections during the campaign and preserved the evidence needed to reproduce, inspect, and re-evaluate the deliverables afterward.