From research assistant to autonomous agent in drug discovery
Agentic AI drug discovery refers to AI systems that not only analyze scientific data or suggest hypotheses but can also plan, coordinate and execute linked experimental and computational tasks across the drug-discovery pipeline with minimal human intervention. After early years of AI acting mainly as a recommendation layer, a new wave of platforms is wiring models into tools, infrastructure and laboratory systems so agents can carry out protocol translation, virtual screening and even lab automation. NVIDIA’s BioNeMo agent toolkit, Medable’s Digital Data Flow Agent, Boltz’s drug-discovery API and cloud-native environments such as Genesis Workbench all point to the same shift: AI agents moving closer to operational control of real-world workflows. Instead of stopping at generating code or slide decks, these systems trigger jobs, update databases and surface ready-to-review results for scientists and clinical teams.
Protocol translation becomes an AI-powered data backbone
In clinical development, protocol setup has long been a drag on study start. Medable’s Digital Data Flow Agent uses AI-powered protocol translation to convert static trial documents into CDISC USDM 4.0, a machine-readable JSON standard that downstream tools can use automatically. Earlier work from the company showed that AI configuration of an eCOA mobile app could cut deployment timelines, which had often stretched 12 to 16 weeks, by about half. The new agent extends that approach, turning protocol design into structured data that flows into eCOA, study documents and operational systems without repeated manual re-keying. According to Andrew Mackinnon at Medable, the goal is to let agents “handle that tactical and administrative burden” so clinical experts can scale their oversight across more sites and studies. This kind of automated data foundation is a critical bridge between trial design and agentic AI drug discovery platforms upstream.
BioNeMo Agent Toolkit gives AI agents scientific “skills” at scale
NVIDIA’s BioNeMo Agent Toolkit targets the heart of autonomous drug screening and design: giving AI agents direct access to specialized scientific models and workflows. The toolkit packages protein-structure prediction, molecular docking, generative chemistry and genomic analysis as documented skills that any compatible AI agent can call. Jensen Huang described the split as “Frontier models are the brains. BioNeMo is the scientific toolbox,” emphasizing that NVIDIA supplies the tools, not the agents themselves. Kimberly Powell highlighted that the platform is harness-agnostic, so developers can plug in their preferred large models while the harness tracks workflow state, tools and rules. With nearly 50 partners, including Eli Lilly, Thermo Fisher Scientific and Dassault Systèmes, the BioNeMo agent toolkit is gaining enterprise validation. It underpins AI agents in pharma research that can move from describing a task like “design a binder” to executing the multi-step domain workflow required to propose and score candidates.
Boltz’s agent-first API lowers the bar for computational screening
On the discovery side, Boltz’s new API was “built for agents as much as for people,” turning its open biomolecular models into an accessible service for autonomous drug screening. The API exposes BoltzProt-1 for protein design and BoltzMol-1 for small-molecule hit discovery, which has been experimentally validated across 10 targets spanning GPCRs, kinases, ion channels and protein–protein interactions. Boltz’s own scientists and engineers already reach these models through coding agents such as Claude Code and Codex; the public API continues that pattern with SDKs and integrations that let AI agents launch screening runs with natural-language prompts or short scripts. According to CEO Gabriele Corso, the backend team designed the system to run thousands of predictions quickly while keeping intellectual property with the user. For teams without extensive GPU infrastructure, this kind of service brings agentic AI drug discovery within reach of smaller biotechs and academic groups.
Toward integrated, cloud-scale agentic workflows in life sciences
Taken together with cloud environments like Genesis Workbench, these tools point toward end-to-end agentic workflows spanning design, simulation and trial execution. A scientist could draft a protocol, have Medable’s Digital Data Flow Agent convert it into CDISC USDM 4.0, and then feed key design parameters into AI agents powered by the BioNeMo agent toolkit or Boltz’s API for target selection and autonomous drug screening. In a cloud workbench, those agents can orchestrate jobs across GPUs, update shared data stores and surface candidates and analyses for human review. The emerging pattern is a layered architecture: AI agents as the control layer, specialized scientific skills beneath, and cloud infrastructure as the execution fabric. As more pharma and biotech partners validate these stacks in production, AI agents in pharma research are likely to move from pilot projects into routine, auditable components of the drug-discovery pipeline.
