ScienceBuddy turns researcher requests, feedback and execution evidence into new tasks and evaluation rubrics. Its inner loop edits the agent harness while holding the model fixed; an outer loop reinforces the model under the improved harness. The authors release the workspace and code with case studies across four scientific task families. The evidence remains a small author-run set rather than an independent productivity study.
ScienceBuddy releases a self-improving scientific-agent workspace
2026-09-16Frontier watchTRIAL
ScienceBuddy turns researcher requests, feedback and execution evidence into new tasks and evaluation rubrics. Its inner loop edits the agent harness while holding the model fixed; an outer loop reinforces the model under the improved harness. The authors release the workspace and code with case studies across four scientific task families. The evidence remains a small author-run set rather than an independent productivity study.
Availability: public product site and GitHub repository
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For teams building research agents, this is a concrete design for learning from normal interactive use without mixing harness changes and model updates into one opaque loop. Trial the trace-to-rubric workflow before adopting recursive model training. Keep human review and held-out evaluations because user feedback can encode local preferences or mistakes.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Loops | inner harness evolution with fixed model; outer reinforcement learning with improved harness |
|---|---|
| Interface | chat, trajectory timeline, compute and result panels |
| Evaluation | case studies across four scientific task families |
| Availability | public product site and GitHub repository |
| Limitation | author-run case studies; no independent productivity or safety evaluation |
Technical impact 4
Engineering value 4
Adoption signal 1
Maturity 2
Verification cost 3
Risk 4