Evidence note: This article rests substantially on a single primary vendor announcement — NVIDIA’s developer blog on running BioNeMo NIM microservices for protein structure prediction in Claude Science. Statements drawn from it describe what NVIDIA itself reports, and are treated here as facts about that announcement rather than as independently corroborated results. Broader statements about how "AI scientists" read papers, form hypotheses, call models, and prioritize experiments are likewise described as reported by that source; independent verification of these capabilities is not established here, and where a specific figure or benchmark would be needed, sources do not provide one.
How agentic AI is changing the way research is done
The framing that agentic AI is shifting how research gets done — moving a language model from a passive text generator to a system that plans steps, invokes external tools, and acts on the results — is one NVIDIA advances in its developer blog on BioNeMo NIM microservices in Claude Science (NVIDIA Developer Blog). The distinction it draws is between an assistant that answers a question and an agent that decides which action to take next and then takes it. Independent evaluation of how far this changes day-to-day scientific practice is not part of that source, so the shift is best read as the direction the vendor describes rather than a measured outcome across labs.
What makes the agentic pattern relevant to science, on this account, is that research is rarely a single query. It is a loop: gather context, form a candidate explanation, test it against a tool or a dataset, and revise. An agent that can carry that loop — rather than handing each step back to a human — is the capability the announcement points to.
AI scientists that can read papers
Reading the literature is the entry point. The value of an agent that can ingest papers is not summarization for its own sake but building the working context an experiment needs: what has already been tried, which methods are standard, and where the open questions sit. That an AI system in this setting can read papers is presented in NVIDIA’s post as part of the agentic research workflow (NVIDIA Developer Blog); the post is the source for the claim, and it is not independently confirmed here.
For a reader evaluating such a capability, the practical questions are the ones the announcement does not settle: how the system handles conflicting findings, whether it distinguishes a preprint from a peer-reviewed result, and how it cites what it relied on. Those remain open and are worth pressing on before treating paper-reading as a solved step rather than a described one.
AI scientists that can propose hypotheses
Beyond reading, the workflow NVIDIA describes includes proposing hypotheses — turning the assembled literature into a candidate explanation or a testable prediction (NVIDIA Developer Blog). This is the step where an agent moves from digesting existing knowledge to generating a new direction to investigate.
The reservation that matters most here is load-bearing: a proposed hypothesis is only as useful as its grounding, and a system that generates plausible-sounding hypotheses is not the same as one that generates correct or novel ones. The source describes the capability; it does not offer a measured hit rate for how often agent-proposed hypotheses hold up, and sources do not confirm one. The claim should be read as "the system can produce hypotheses," not "the hypotheses are reliably good."
AI scientists that can call models
The step that separates an agent from a chatbot is the ability to call models and tools directly. In this workflow, that means invoking specialized scientific models — not asking the language model to imitate their output, but routing a task to a purpose-built model and using what comes back. NVIDIA’s post frames the agent’s ability to call models as a core part of the pattern, with BioNeMo NIM microservices as the models being called (NVIDIA Developer Blog).
This is the practical hinge of the whole approach. A general model reasoning about protein structure in prose is a different thing from a general model that hands the structure-prediction task to a dedicated model and reasons over the actual prediction. The latter is what the microservice packaging is meant to enable.
AI scientists that determine which experiments to prioritize next
The final loop-closing capability the announcement describes is prioritization — the agent deciding, given what it has read, hypothesized, and computed, which experiment is worth running next (NVIDIA Developer Blog). This is what turns a set of individual tool calls into something resembling a research strategy: not just running an analysis, but choosing which analysis is most informative to run.
How well an agent prioritizes is precisely the kind of claim that needs measurement rather than description, and the source provides the capability, not an evaluation of its judgment. Whether agent-chosen experiment ordering outperforms a researcher’s own remains open.
Running NVIDIA BioNeMo NIM microservices for protein structure prediction in Claude Science
The concrete instance NVIDIA documents is running its BioNeMo NIM microservices for protein structure prediction inside Claude Science (NVIDIA Developer Blog). NIM (NVIDIA Inference Microservices) is the packaging layer: a scientific model is exposed as a callable service, which is what lets an agent invoke it as one tool among others rather than embedding the model directly. Protein structure prediction — inferring a protein’s three-dimensional shape — is the task chosen to demonstrate the pattern, because it is a case where a general language model cannot substitute for a specialized model and the value of calling out is clear.
Put together, the pieces map onto the capabilities above: the agent reads the relevant biology, frames a structural question, calls the BioNeMo microservice to predict the structure, and uses the result to decide what to look at next. That end-to-end description is NVIDIA’s own; it establishes that the integration exists and how it is intended to work, not how accurate the predictions are or how the setup performs against alternatives. No performance figures or benchmark results are drawn here, because the cited source is a primary vendor announcement and independent numbers are not available to report.
For a practitioner, the actionable takeaways are modest but real: the microservice model means a structure-prediction step can be wired into an agent as a callable service, and the Claude Science integration is the documented path for doing so. Anyone planning to rely on it should treat the accuracy, cost, and comparative performance of that path as questions still to be measured rather than settled by the announcement.