AI scientists are becoming a new interface for scientific computing. These agents read papers, write code, generate hypotheses, call APIs, and inspect files. But science is not software engineering. No test suite turns green when a hypothesis is correct. Discovery stays iterative, uncertain, and grounded in the physical world. That gap is what NVIDIA is targeting with the BioNeMo Agent Toolkit, an open-source repository of “skills” that turn NVIDIA’s biomolecular models into tools AI agents can call reliably. Published as a hands-on walkthrough on the NVIDIA Developer Blog, the toolkit aims to solve a fundamental problem: a general coding agent pointed at biology will not produce new medicines. In biomolecular research, an agent’s ceiling is set by the tools it can use correctly, efficiently, and without guesswork.
What Is the BioNeMo Agent Toolkit?
The BioNeMo Agent Toolkit is an open-source repository of documented, callable skills for AI agents operating in drug discovery and life sciences. Each skill packages an NVIDIA biomolecular model — such as protein folding, molecular docking, generative chemistry, genomics analysis, protein design, or biomarker discovery — into a tool an agent can discover, select, invoke, and interpret. The toolkit is built on two layers: an accelerated tool layer powered by NVIDIA NIM (NVIDIA Inference Microservices) and BioNeMo open models, accelerated by libraries like cuEquivariance for structure models and Parabricks for genomics; and agent-ready interfaces that package each capability as a skill. Agents read a skill’s documentation, understand its inputs, parameters, expected artifacts, and failure modes, then act autonomously.
TL;DR
- BioNeMo Agent Toolkit packages NVIDIA biomolecular models as documented, callable agent skills.
- Skills span protein folding, docking, generative chemistry, genomics, and protein design.
- NVIDIA reports task completion rising from 57.1% to 100% when agents use skills.
- Agents averaged 2x more passing assertions per 1,000 tokens with skills.
- Hosted NIM endpoints suit quick access; local NIM suits repeated iteration and data locality.
How a BioNeMo Skill Works
Every skill is a directory containing a SKILL.md file. That file includes YAML frontmatter plus instructions, optional references, and optional scripts. An agent reads it as documentation and then acts on it. The prompt pattern stays consistent across models. For example, using the OpenFold3 skill, the agent can be instructed: “Use the OpenFold3 BioNeMo Skill to fold MKTVRQERLKSIVR with the NVIDIA API endpoint at https://build.nvidia.com/openfold3”. The same shape applies to other NIMs for biology, including Boltz-2, DiffDock, GenMol, ProteinMPNN, RFdiffusion, and Evo 2.
Installation is handled through the open-source skills CLI. A user can browse and pick a skill interactively with npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit, or install a single skill for a specific agent with npx skills add NVIDIA-BioNeMo/bionemo-agent-toolkit –skill boltz2-nim –agent claude-code. Deployment is a choice: use hosted NIM endpoints for fast access without managing infrastructure, or move selected models local when lower warm latency, data locality, or repeated iteration is needed. The Model Context Protocol (MCP) server wrappers expose open models not yet packaged as NIM, ensuring broader coverage.
Benchmark: The Measured Impact of Skills
NVIDIA measured whether skills actually improve an agent’s loop. All reported metrics came from Codex CLI running GPT-5.5 fast. The team compared the same agent with and without each skill.
Task completion was the first metric. Without skills, the agent completed 57.1% of required tasks on average. With access to NIM skills, completion reached 100% across all tested scenarios.
Efficiency was the second metric. NVIDIA counted passing assertions — the individual steps that compose a task. With skills, an agent produced 2x more passing assertions per 1,000 tokens. That gain held consistently across all ten NIM skills evaluated.
Use Cases With Examples
- Protein structure prediction: An agent folds a peptide sequence with Boltz-2 or OpenFold3 and returns a CIF file for downstream inspection.
- Multiple sequence alignment: An agent generates an MSA with MMseqs2 through the MSA Search skill, outputting an A3M file.
- Generative chemistry: An agent generates candidate molecules with GenMol, outputting SDF or SMILES for filtering.
- Protein binder design: The generative_protein_binder_design workflow chains three models — RFdiffusion builds a backbone, ProteinMPNN designs the sequence, and OpenFold3 validates the fold — in a single meta-skill.
- Each loop follows the same shape: the agent selects a model, prepares inputs, runs it, inspects outputs, and explains results with scientific caveats.
Agent With Skills vs. Without Skills: A Comparison
- Task completion: General agent (no skills) 57.1% average vs. Agent + BioNeMo Skills 100% average.
- Token efficiency: Baseline vs. 2x passing assertions per 1k tokens.
- Model selection: Guesses tool, format, and inputs vs. Reads purpose, inputs, and artifacts.
- Deployment: Manual setup from source vs. Hosted or local NIM, documented.
- Failure handling: Unknown failure modes vs. Documented failure modes per skill.
- Workflows: Isolated single calls vs. Multi-step meta-skills like binder design.
Getting Started with BioNeMo Skills
The prerequisites are minimal. You need an agent runtime such as Claude or Codex, and an NVIDIA API key for hosted BioNeMo NIM endpoints. A GPU node is optional for local NIM deployment. NVIDIA recommends pointing the agent at the repository first, letting it enumerate the available capabilities before it acts. Then hand it a single skill to operate one model.
Two cautions are flagged. The build.nvidia.com endpoints are for small-scale development and testing only — not production-grade inference. Additionally, NVIDIA stresses validation: check low-confidence structures and filter generated molecules before trusting them.
What This Means for Drug Discovery Teams
The BioNeMo Agent Toolkit turns a general-purpose AI coding agent into a focused scientific assistant that can reliably execute complex biomolecular workflows. For teams in pharmaceutical research, computational biology, or AI-driven drug discovery, the implications are immediate: the ability to chain models like RFdiffusion, ProteinMPNN, and OpenFold3 without manual integration reduces iteration time from days to minutes. The open-source nature of the toolkit means any organization with an NVIDIA API key can begin testing today. The next step is to deploy a skill for one model — such as OpenFold3 for structure prediction — and measure how it changes your own agent’s completion rate and token efficiency. The toolkit is available on GitHub, and the technical details are published on the NVIDIA Developer Blog. For teams serious about building AI scientists, this is a practical starting point.
