Nvidia has launched the BioNeMo Agent Toolkit, a set of tools designed to help AI agents carry out life science tasks more reliably. The company says the toolkit is meant to make biomolecular models easier for agents to discover, select and use in research workflows that involve proteins, molecules and genomics.
The announcement frames AI scientists as a new kind of interface for scientific computing, capable of reading papers, writing code, generating hypotheses and calling tools. But Nvidia argues that biology creates a different challenge from software engineering, where tests can verify whether code works. In life science discovery, models need to be chosen carefully, inputs must be formatted correctly and outputs must be interpreted with scientific caution.
BioNeMo is Nvidia’s platform for that process. It combines accelerated model services with agent-ready wrappers the company calls BioNeMo Skills. These skills package biomolecular capabilities such as structure prediction, molecular generation, docking, sequence analysis, alignment and genomics into callable services. Nvidia says the platform is built on its accelerated digital biology stack, including NIM and BioNeMo open models, and uses libraries such as cuEquivariance for structure models and Parabricks for genomics.
According to the company, the skills are designed to give agents more than a simple endpoint to call. Each one documents the model’s purpose, required inputs, optional parameters, expected outputs and possible failure modes. Nvidia says that context allows an agent to determine which tool is appropriate, prepare a valid request and interpret the resulting files, such as CIF, SDF, FASTA, A3M or SMILES artifacts.
The toolkit also supports Model Context Protocol server wrappers for open models that are not yet packaged as NIM services. That approach is intended to expose those models through the same agent-friendly pattern, regardless of the backend runtime.
Nvidia says teams can use the toolkit with an agent runtime such as Claude or Codex. The company recommends starting by pointing an agent to the BioNeMo Agent Toolkit repository so it can discover available capabilities before acting on them. From there, the agent can use a specific skill or MCP wrapper to learn what the model does, when to use it and what output to expect.
Deployment flexibility is another key part of the launch. Nvidia says BioNeMo models can run as hosted endpoints or on local infrastructure. Hosted NIM endpoints are positioned as the easier option for evaluation, occasional use or workflows that depend on infrastructure-heavy services. Local deployment is aimed at iterative tasks that require repeated calls, lower latency, data locality or tighter control.
The blog post includes examples of workflows an AI scientist might perform, including generating a multiple sequence alignment with MMseqs2, folding a peptide with Boltz-2 or OpenFold3, generating molecules with GenMol and docking ligands with DiffDock. Nvidia says the toolkit is intended to help an agent move through those steps as part of a research loop that includes selecting a model, preparing inputs, running inference, inspecting the result and explaining the outcome.
Nvidia also says its internal benchmarking found that access to BioNeMo Skills improved agent task completion rates from 57.1 percent to 100 percent in the tested setup. The company reported that the skills roughly doubled token efficiency as well, measured by the number of passing assertions per 1,000 tokens.
The broader pitch is that BioNeMo turns biomolecular models into production-ready tools for agents rather than standalone models. Nvidia says that should make iterative research workflows faster and more reliable, especially as life science teams experiment with AI systems that can do more than answer questions.