From descriptive chatbots to agents that run drug discovery
AI agents in drug discovery are software systems that combine large language models with domain-specific tools so they can translate natural-language scientific goals into concrete, multi-step computational workflows that execute, track, and refine tasks such as target analysis, virtual screening, and protein or molecule design across regulated life sciences environments.
The important shift is that AI agents in drug discovery are no longer confined to describing what a scientist could do. They now call specialized tools, run screens, and return ranked candidates. This move from theory to execution is visible in two complementary directions: a drug screening API from Boltz built “for agents as much as for people”, and the BioNeMo agent toolkit, which turns NVIDIA’s life-sciences stack into callable skills for any agent. Together they mark a clear opinionated inflection point: if you are still treating AI as a fancy search bar, you are already behind. The frontier is autonomous life sciences workflows that are cheap, fast and tightly governed.
Boltz’s drug screening API: a natural-language pipeline for agents
Boltz’s drug-discovery API is an explicit bet that AI agents should be first-class users of scientific infrastructure. The company has wrapped its BoltzProt-1 protein design pipeline and BoltzMol-1 small-molecule hit-discovery pipeline in a usage-priced drug screening API, then wired that interface into coding agents like Claude Code, Codex and Gemini CLI. In practice, that means a chemist—or an AI agent acting for them—can request a hit-discovery run in plain English and have the agent handle authentication, configuration and execution end to end.
In a live test, an AI agent using Claude Code called the Boltz API to screen a handful of commercially available compounds against EGFR kinase, a classic cancer target, through a natural-language prompt. The agent installed and authenticated the command-line tool on its own, refused to invent protein or compound data, estimated the cost per molecule before spending, and returned ranked structures in about 3.5 minutes. Experimental validation of BoltzMol-1 across 10 targets, spanning GPCRs, kinases, ion channels and protein–protein interactions, underlines that this is not a toy demo but a pipeline aimed at real hit discovery.
Why the BioNeMo agent toolkit matters more than another model
Where Boltz offers a single drug screening API endpoint, the BioNeMo agent toolkit aims to be the standard toolbox for autonomous life sciences workflows. NVIDIA describes it as an open, harness-agnostic platform that gives AI agents or software platforms the building blocks to specialize for science, rather than another monolithic model. It packages life-sciences software and models for protein-structure prediction, molecular docking, generative chemistry and genomic analysis as documented skills an agent can call independently.
This design choice is opinionated: a general-purpose model “doesn’t understand” the five to seven domain-specific tasks hidden inside a request like “design me a binder,” so BioNeMo makes those tasks explicit and callable. The toolkit is agent-agnostic, so any developer, regardless of harness, model or platform, can access the same accelerated scientific tools with customizable governance. Debuting with adoption from nearly 50 partners, including Eli Lilly, Thermo Fisher Scientific and Dassault Systèmes, it signals that enterprise players are willing to trust agent-driven automation for complex scientific workflows—provided the tools are validated and the rules of use are encoded.
Why this shift is happening now—and why it is still contested
This agentic turn did not appear in isolation. Since AlphaFold 2 cracked protein-structure prediction in 2020 and 2021, an ecosystem of biomolecular models such as Boltz, Chai, OpenFold and Protenix has grown rapidly. Yet many drug developers still have to choose between running open models on their own GPU clusters or renting access from workflow providers like Tamarind Bio, Rowan or BioNeMo. Toolkits and APIs for AI agents are a pragmatic answer: encapsulate these powerful models in governed interfaces that are callable, auditable and affordable.
At the same time, scientist trust in AI is uneven. One recent study of more than 25,000 agent runs found that LLM-based agents can execute workflows but ignore gathered evidence in most cases and rarely revise conclusions when data conflicts with earlier assumptions. Surveys show that only a small share of life-sciences leaders see strong value from AI in wet labs or end-to-end scientific workflows so far. Yet demand keeps bubbling up through “shadow AI”, with many scientists using public generative tools through personal accounts despite security and compliance risks. The message: the need for automation is real, but so is the skepticism.
From experimental agents to embedded “little experts”
The direction of travel is clear: AI agents in drug discovery are moving from experiments on the side to embedded utilities. Boltz reports that agent plugins, including integrations with Anthropic’s Claude and OpenAI’s Codex, have become the main way its own chemists and protein engineers interact with its models. On the BioNeMo side, four AI agents already run across R&D stages on the toolkit, with future progress measured in better task completion, fewer failed tool calls, lower integration overhead and faster movement from raw data to prioritized candidates.
We should expect these “little experts” to slide inside electronic lab notebooks and scientific data systems, quietly handling routine but complex steps—from configuring a virtual screen to triaging hits—under human oversight. The next challenge is cultural, not technical: winning over scientists who have seen LLMs hallucinate and overclaim. If platforms like Boltz’s drug screening API and the BioNeMo agent toolkit keep focusing on validated tools, clear governance and measurable outcomes, they stand a real chance of turning AI agents from flashy demos into reliable lab colleagues.
