What they did
Bioinfoysis represents each user request as a persistent analysis run rather than a sequence of disposable model interactions. A planner maintains an executable checklist and revises unfinished steps after workers return structured handoffs linking each intermediate result to its agent, checklist step, and version of the plan.
The system also uses role-specific context, persistent memory, governed bioinformatics skills, and a controlled runtime that validates generated scripts, tables, and figures before downstream use. The authors evaluated it on BixBench and the SeqQA2 and DbQA2 question-answering tracks of LAB-Bench 2, testing the harness with four underlying language models.
Key findings
- Bioinfoysis achieved 82.4% accuracy on BixBench, which the authors report as state of the art.
- Across four language models, average SeqQA2 accuracy rose from 27.81% without the harness to 64.13% with it, a gain of 36.32 percentage points.
- Average DbQA2 accuracy increased from 3.13% to 31.25%, a gain of 28.12 percentage points.
- The reported results indicate that persistent planning, artifact validation, and controlled evidence flow can improve performance independently of the underlying model choice.
Why it matters
Bioinformatics conclusions often depend on long chains of data processing, code execution, and intermediate interpretation. By preserving provenance and preventing results from obsolete plans from being silently reused, Bioinfoysis offers a framework for making agentic analyses more traceable and less dependent on a model remembering an entire workflow correctly.
Caveats
The abstract does not provide benchmark sample sizes, uncertainty estimates, per-model results, error categories, or ablations isolating the contribution of each harness component. Benchmark accuracy also does not establish that the system is reliable for open-ended or safety-critical biological research, where data quality, reproducibility, and expert review remain essential.