ScienceBuddy is an interactive scientific research workspace that embeds continuously improving scientific agents into researchers' everyday workflows. The system converts researchers' requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning.
The core mechanism is recursive‑in‑recursive self‑improvement: an inner recursion refines the task harness while keeping the model fixed, and an outer recursion trains the model under the improved harness. Harness evolution shapes the training experience, and model learning creates new opportunities for harness adaptation.
We present case studies of researcher interaction, harness refinement, and model learning across four scientific task families. By releasing ScienceBuddy as a research product, we aim to advance discovery intelligence—enabling scientific AI to evolve through sustained collaboration with researchers and alongside the research it supports.
Review: The work offers a reproducible framework that demonstrates bidirectional iteration between model and task harness, laying groundwork for future adaptive scientific systems.