Build AI Agents for Life Sciences with NVIDIA BioNeMo and Nebius in partnership with J.P. Morgan

Skills provide instructions. Tool access provides the ability to execute those instructions. An agent may know how a scientific model works but still need a running service, an endpoint and the appropriate credentials to use it.

René Schönfelder explains this distinction in the Nebius demo. Skills describe how to use the models, while APIs and Model Context Protocol (MCP) servers connect the agent to the services that perform the computations. Both parts are needed for an executable workflow.

Protein binder design involves several steps: understanding a target, generating candidate designs, predicting structures and comparing candidates. An agent can coordinate specialized tools across these steps and use their outputs to decide what to run next.

NVIDIA presents an example that combines structure prediction, sequence alignment and binder generation. The agent organizes a computational research workflow. Its generated candidates and predicted scores still need scientific assessment before researchers can draw conclusions about their effectiveness.

The demonstration uses an OpenClaw-based agent hosted through Nebius Serverless AI. It connects the agent to a reasoning model, scientific model services and optional Tavily web search.

René shows configuration for authentication and service credentials, then walks through example research tasks and their outputs. Nebius Token Factory supplies a hosted reasoning-model option, while the scientific computations run through separate model services. This separates the agent’s coordination work from the compute required by individual scientific tools.

Hosted endpoints can simplify initial experiments because teams do not need to deploy each model themselves. Self-hosting provides more control over the model environment, infrastructure and region where computations run.

The workshop discusses deploying NVIDIA NIM services on GPU instances and configuring the agent to access them. René also describes hosting the agent, reasoning model and scientific services within a team’s own resources. The appropriate setup depends on the workflow and its data-handling and operational requirements.

Evaluation should reflect the specific research task. An agent for molecular design needs different checks from an agent for literature analysis, lab automation or medical imaging. There is no single evaluation that establishes scientific accuracy across all these applications.

In the Q&A, Mahan Salehi recommends starting with the processes a team already uses to judge successful research outcomes and turning them into evaluations. Correct tool execution is one part of that assessment; teams also need to evaluate whether the resulting outputs meet their scientific criteria.

Costs can come from several components: hosting the agent, using a reasoning model and running scientific tools. A small CPU instance may handle agent coordination while GPU-backed services perform more demanding model computations.

The workshop distinguishes token-based reasoning-model charges from CPU or GPU infrastructure charges. NVIDIA also explains how skills can reduce wasted attempts and token use by giving the agent relevant tool instructions. The cost of hosting the agent alone does not represent the total cost of a research workflow.