OpenAI4S: Code as Action, Science as Sessions

An open-source agent harness for computational science from Li Yuan's PKU-YuanGroup at Peking University's Shenzhen Graduate School, launched by the Peking University–YuanKong Intelligence AI Joint Research Laboratory. Its premise is that a long-running study has to stay inspectable, resumable and reproducible, so the harness is built around a persistent computing runtime rather than a chat loop. One agent loop drives two channels: provider-native JSON tool calls handle orchestration (metadata, services, permissions and workflow management), while each scientific action is a complete code cell run in a persistent Python or R kernel (code as action). Each session is kept as a research record (science as sessions): an append-only Action Ledger, per-cell execution records, versioned artifacts, environment records and workspace checkpoints support recovery, branching and export. Operating-system sandboxing, permission mediation, and screening of code and trajectories for selected biological and chemical risks act as safeguards. Domain knowledge comes from bundled Skills, code recipes the model retrieves on demand: 604 at the time of the paper (43 curated plus 561 from the MIT-licensed bioSkills collection). The engine is provider-neutral, with adapters for the OpenAI, Anthropic and Gemini API formats.

The paper evaluates it on 36 research scenarios across six tasks: retrosynthesis, molecular dynamics, protein binder design, protein mutation, catalyst structure–activity screening and mineral spectra analysis. A deterministic-first evaluator, in which programmatic checks take precedence over a constrained LLM judge, scores each generated research repository on accuracy (40%), workflow completeness (35%) and reproducibility (25%). OpenAI4S averages 7.83 out of 10, against 5.79, 6.36 and 6.12 for Claude Code driven by Kimi-K3, GLM-5.2 and Claude Opus 4.8, and 8.70 for a reference workflow built from human-designed pipelines. It ranks first on protein binder design, molecular dynamics and catalyst screening and second on the other three, with its largest gains on molecular dynamics and protein binder design, the longest and most compute-heavy workflows. The authors caution that these scores compare complete systems (model, harness, prompts and tools) rather than matched ablations, that the margin over GLM-5.2 shrinks to about 0.04 points without the protein tasks, and that environment specification and full rerunnability remain weak for every system tested, their own included. The code was open-sourced on July 6, 2026 (the date used here), with releases through v0.3.0 on September 17 and a PyPI package; the paper was posted on September 14. The repository is MIT-licensed and had about 600 GitHub stars at filing.