NVIDIA Announces BioNeMo Agent Toolkit with traction from nearly 50 partners, including Lilly, Thermo Fisher and Dassault

NVIDIA has unveiled the BioNeMo Agent Toolkit, an open, harness-agnostic platform that gives AI agents or software platforms the building blocks to specialize for science. Debuting with adoption from nearly 50 partners, including Eli Lilly, Thermo Fisher Scientific and Dassault Systèmes, the toolkit packages NVIDIA’s life-sciences software, models for tasks such as protein-structure prediction, molecular docking, generative chemistry and genomic analysis, as documented “skills” an agent can call on its own, so a general-purpose AI agent can carry out real scientific work rather than just describe it. “Frontier models are the brains. BioNeMo is the scientific toolbox,” said Jensen Huang, founder and CEO of NVIDIA.

A flexible and specialized framework

In a press conference, NVIDIA Vice President of Healthcare Kimberly Powell highlighted four AI agents built on the platform across R&D stages. “Because it’s agent-agnostic, any developer, regardless of which harness, model, or platform they build on, can access the same accelerated scientific tools and skills with customizable governance,” she said. “Think of the harness as the operating system for an agent. It connects the models to the right tools, keeps track of where the workflow is, remembers what happened in earlier steps, and enforces the rules about what the agent can and cannot do.”

Powell drew a line around what NVIDIA itself is offering. “We’re not even building agents. That’s why the product name is the NVIDIA BioNeMo Agent Toolkit. These are the tools we provide to agents, and it’s agent-agnostic.”

A general-purpose model cannot break a scientific goal into its component steps on its own, Powell argued. “They don’t understand what it means when you say ‘go design me a binder,’ the five, six, seven individual domain-specialized tasks that it takes. We need to create the conditions for agents to specialize, because the work in science is deeply specialized.”

Part of a growing landscape

The NVIDIA news is part of a broader trend of maturing AI in life sciences. Boltz, an open biomolecular-model lab, last week released BoltzMol-1 for small-molecule hit discovery and BoltzProt-1 for protein design behind a usage-priced API, turning its open models into a metered inference service. Alongside that launch, Boltz said agent plugins, including integrations with Anthropic’s Claude and OpenAI’s Codex, have become the main way its own chemists and protein engineers run its models. Days earlier, OpenAI, one of the frontier labs NVIDIA names as integrating the toolkit, and Molecule.one reported a “near-autonomous AI chemist” that paired GPT-5.4 with Molecule.one’s Maria agent and an automated lab to run more than 10,000 reactions and improve a stubborn sulfonamide coupling, gains human chemists then reproduced at the bench. In February, OpenAI and Ginkgo Bioworks announced progress in using GPT-5 for cell-free protein synthesis.

At a launch event for BoltzMol-1 and BoltzProt-1, Gabriele Corso, Ph.D., CEO at Boltz, noted that most of its scientists, chemists and protein engineers have been interacting with its models via coding agents like Claude Code, Codex and Gemini. “It’s very exciting to see this movie in public,” he said.

Growing AI traction in science, uneven trust

Meanwhile, scientists’ relationship with AI looks more complicated than steady adoption suggests, and in some quarters distrust runs high. In April, a preprint titled “AI scientists produce results without reasoning scientifically” reported that across more than 25,000 agent runs in eight scientific domains, LLM-based agents executed workflows competently. The paper also highlighted that the agents ignored evidence they had already gathered in 68% of cases and rarely revised their conclusions when data contradicted them.

A Pistoia Alliance poll of 300 life-sciences leaders that same month found only 1% saw AI delivering value in the wet lab, and 13% in automating scientific workflows and experiments. The clearest gains were in text-heavy work, searching literature and writing reports.

That conflicted relationship shows up in broader sentiment data, too. A June 2026 Nature poll of 1,907 scientists found nearly half felt broadly negative toward AI, and 63% said the risks of using large language models for data analysis and the scientific literature outweighed the benefits; just 23% saw a positive impact on their research. Meanwhile, Wiley’s 2025 ExplanAItions survey of more than 2,400 researchers showed the same push and pull. AI adoption climbed to 84% even as worry about hallucinations rose from 51% to 64%, and the share who believed AI already outperforms humans on important tasks fell from more than half to under a third. Yet demand keeps surfacing anyway. A Sapio Sciences survey of 150 scientists, published in January, found 45% were already using public generative-AI tools through personal accounts, a “shadow AI” habit the company called “pretty alarming” given the security, IP and compliance risks.

In an interview with R&D World, Christian Baber, chief portfolio officer for the Pistoia Alliance, traces that gap to a question of agency, “human in the loop versus AI in the loop versus nothing in the loop.” Scientists won’t cede control entirely, he said, and can’t when it comes to regulatory submissions. He does see the agentic shift coming, with “little experts” taking routine tasks and getting embedded into electronic lab notebooks, but acknowledges the task of winning over science’s trust will take work.

Asked what would convince skeptics, Powell said the answer is giving agents “validated domain-specific scientific tools that researchers already use,” packaged so the agent knows each tool’s purpose, inputs, outputs and troubleshooting. The goal, she said, is measurable progress: better task completion, fewer failed tool calls, less integration overhead, faster movement from data to prioritized candidates.

Baber frames the larger stakes in broad strokes. “You can do things better, or you can do better things,” he said. “At the moment, we’re seeing AI being used to do things better. The real revolution comes when you’re doing better things.” By that test, today’s AI sits in the first column. Whether tooling like NVIDIA’s moves the field into the second is the question this launch leaves open.