The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a setup script (scripts/setup_environment.py) that uses the subprocess module to manage environment configuration.
It executes conda commands to create and verify the execution environment.
It runs a temporary test script using the current Python executable to verify the installation.
[DYNAMIC_EXECUTION]: The core functionality of the Biomni agent (A1.go()) involves autonomous code generation and execution.
The documentation in SKILL.md and references/api_reference.md explicitly states that the agent generates and executes Python code to perform analysis.
Users are warned that code is executed with full system privileges and are advised to use isolated environments like Docker for production use.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process a wide variety of external biomedical data formats, which represents an indirect injection surface.
Ingestion points: Data is ingested via file paths provided in natural language queries to agent.go(), supporting formats like H5AD (single-cell RNA-seq), VCF (variants), CSV, and GWAS summary statistics.
Boundary markers: The skill relies on standard LLM prompting; specific boundary markers for untrusted file content are not explicitly detailed in the instructions.
Capability inventory: The agent has the capability to execute arbitrary Python code, write files (reports), and access the network for database queries.
Sanitization: While the framework manages the high-level workflow, the generated code executes with the privileges of the host user unless manually sandboxed.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of a large (~11GB) biomedical data lake upon first use.
Data is fetched from established and trusted scientific organizations including Ensembl, NCBI, UniProt, and various biomedical ontologies.
Documentation provides links to official repositories and papers hosted on trusted platforms like GitHub (snap-stanford), bioRxiv, and Hugging Face.

biomni