The late nights at Ginkgo Bioworks used to blur together. Scientists hunched over cell-painting assays, wrangling transcriptomic data, waiting weeks for results that might—if the protocols held—be publication-ready. Then last year, something changed. The same analyses started running themselves, orchestrated by an AI agent that compressed what had been month-long slogs into a matter of hours. The quality? Still publication-grade, the researchers confirmed. But the work itself had migrated from human hands to software that plans, executes, and iterates on its own.
They're calling it agentic biology, and two years ago the category barely existed.
Now it may represent the most disruptive force in drug discovery since the era of high-throughput screening reshaped the pharmaceutical industry. Unlike earlier AI systems—which required humans to frame every query, interpret every output, hold every metaphorical hand—these new platforms chart their own experimental paths. They navigate databases. Write protocols. Loop back to refine hypotheses when the data don't cooperate. Less like tools, more like tireless research assistants who happen never to clock out.
The market is taking notice. McKinsey's analysts estimate generative AI could unlock somewhere between $60 billion and $110 billion in annual value for pharma and medical products companies, with agentic systems claiming a growing slice of that windfall.
The Numbers Don't Lie, But They Don't Tell the Whole Story
By the numbers alone, the trajectory looks almost inevitable. AI in drug discovery commanded a $2.35 billion market in 2025, according to Grand View Research. By 2033, analysts project it will hit $13.77 billion—a compound annual growth rate hovering near 25%. Lab automation, the hardware backbone that agentic AI increasingly commandeers, is tracking toward $11.9 billion by 2029. Federal dollars are flowing, too. The National Science Foundation committed $75 million to U.S. BioFoundries last August, a bet that autonomous, AI-driven biology infrastructure can accelerate everything from synthetic biology to advanced materials.
Yet the deeper transformation isn't really about dollar figures. It's architectural.
Traditional AI in biology has been narrow by design—predict this protein structure, screen that compound library, flag a suspect gene variant. Agentic AI operates differently. It strings together entire workflows, ingesting literature, designing experiments, translating natural language into machine-executable protocols, interpreting whatever results emerge, then proposing the next round of tests. The system doesn't wait for instructions. It reasons through research problems end-to-end.
"AI-native biology requires an integrated environment that unifies agent architectures with the tools biologists actually use," says Kexin Huang, CEO and co-founder of Phylo, a South San Francisco startup that launched its Biomni Lab platform this past February. The company's agentic environment wrangles more than 300 databases and tools—everything from COSMIC cancer genomics to Addgene plasmid repositories—letting scientists interact in plain English. Phylo claims its alpha users are firing off thousands of queries monthly, which suggests adoption may be moving past curiosity into operational dependence.
Phylo has company. Benchling, the life sciences data platform that's become infrastructure for over 1,300 companies and 7,500 academic organizations, introduced embedded agents last October. The agents—Compose, Data Entry, Deep Research—sit inside electronic lab notebooks and LIMS systems, automating literature synthesis, data entry drudgework, and protocol generation. Meanwhile, Isomorphic Labs, Alphabet's drug discovery arm, raised $600 million in March 2025 and deepened partnerships with Novartis and Johnson & Johnson, building on AlphaFold 3's ability to predict molecular complexes across proteins, nucleic acids, and ligands.
None of this happened overnight. But the pace is accelerating in ways that feel, to some longtime observers, almost destabilizing.
The Convergence That Made This Possible
Several forces had to align before agentic biology became viable. Foundation models for biology matured fast between 2023 and 2025, maybe faster than anyone expected. AlphaFold 3, published in Nature in May 2024, extended structure prediction to multi-component assemblies—the exact scenarios drug designers lose sleep over. EvolutionaryScale's ESM3, released a month later, went further: it's a generative protein model that designs novel sequences from scratch. The company demonstrated this with esmGFP, a fluorescent protein the model invented that shares less than 60% sequence identity with any known GFP variant.
These models don't just predict. They imagine. And when you pair generative biology with agentic orchestration—systems that autonomously plan, execute, and iterate—you get something that starts to resemble a research colleague more than a research instrument. A 2025 survey by BCG and MIT Sloan Management Review found that 76% of respondents now view agentic AI as a coworker rather than a tool. That's not just semantic hairsplitting; it reflects a genuine expansion of the technology's decision-making latitude.
The infrastructure caught up, too. Autoprotocol and Synthace Antha provide machine-readable standards for automating liquid handlers, plate readers, robotic systems. Multi-agent architectures, documented in recent arXiv preprints, can translate natural language goals into executable protocols and self-correct when errors pop up—a crucial capability if you want reliability in wet-lab settings. Recursion Pharmaceuticals partnered with NVIDIA to train foundation models on high-dimensional phenomics data, distributing them via NVIDIA's BioNeMo platform. The result is a rapidly maturing stack: models that understand biology, orchestration layers that translate intent into action, and hardware that executes without someone hovering over it.
Regulatory shifts matter, though they're uneven and sometimes contradictory. The FDA published draft guidance in January 2025 on using AI to support regulatory decision-making for drugs and biologicals, followed by "Guiding Principles of Good AI Practice in Drug Development" a year later. The European Medicines Agency released its reflection paper on AI in the medicinal product lifecycle in September 2024, signaling acceptance with risk-based guardrails rather than outright prohibitions. In the U.S., President Trump rescinded a sweeping AI safety executive order last January, though sector-specific regulators like the FDA continue developing domain guidance. The message to industry: deploy thoughtfully, validate rigorously, but don't sit on your hands waiting for perfect clarity that may never arrive.
From Academic Experiment to Commercial Infrastructure
Phylo's path illustrates how quickly academic research can scale into commercial infrastructure—when the timing's right. The company's roots trace to Stanford, where co-founder Yuanhao Qu built AI agents to automate his own biology workflows because, well, he got tired of doing it manually. That work evolved into Biomni, an open-source agentic platform published in a preprint mid-2025. The preprint documented an "action discovery agent" that mined tens of thousands of papers across 25 biomedical domains—gene prioritization, drug repurposing, rare disease diagnosis, molecular cloning—to construct a unified agentic environment. The generalist agent integrates language model reasoning, retrieval-augmented planning, and code-based execution.
By February 2026, Phylo commercialized that research into Biomni Lab, closing a $13.5 million seed round co-led by Andreessen Horowitz and Menlo Ventures' Anthology Fund (a collaboration with Anthropic). Jorge Conde of a16z called the adoption signal—more than 7,000 labs already using the open-source platform—"a rare academic-to-real-world transition that actually stuck." Matt Kraning at Menlo framed it as bringing "AI-native productivity to the lab," though that phrase has become something of a cliché in venture pitches lately. Still, Phylo claims Biomni Lab delivers more than 20% performance improvement over existing agent systems on standard benchmarks and long-horizon evaluations.
The Ginkgo case study, sparse as the public details are, hints at practical impact. Researchers there used Biomni Lab to accelerate complex omics analyses that typically consumed weeks, compressing them into hours while maintaining publication-quality rigor. It's one data point. But it suggests agentic systems can handle not just routine queries but multi-step, data-intensive workflows where human bottlenecks have historically mattered most.
Elsewhere, Isomorphic Labs is constructing what amounts to a design engine powered by AlphaFold 3, orchestrating it into drug discovery partnerships with pharmaceutical giants. The $600 million it raised last March—valuing the company's approach of combining predictive biology with iterative screening—was enough to attract serious institutional capital. Recursion, meanwhile, trains foundation models on phenomics data and exposes them via NVIDIA's infrastructure, aiming to map cellular responses at scale and mine them for therapeutic insights.
On the automation side, researchers published a self-driving protein engineering platform called SAMPLE in Nature Chemical Engineering. The system uses robotic hardware and agent planners to autonomously design experiments, execute them, and refine designs to hunt down thermostable enzymes. Another group demonstrated multi-agent control of liquid handlers, translating natural language into protocols with built-in self-correction. These aren't lab curiosities anymore—they're production-grade systems closing the design-test-learn loop without constant human intervention.
The Gap Between Pilot and Production
If the past two years established proof of concept, the next two will test whether any of this scales. Or endures.
McKinsey's analysis is characteristically blunt: generative AI in pharma could unlock that $60 billion to $110 billion windfall annually, yet as of 2025, only 5% of companies have realized sustained competitive differentiation at scale. The chasm between pilot projects and production systems remains wide, perhaps wider than the hype cycle suggests. Data plumbing, risk management, change management—the decidedly unglamorous challenges—are gating factors. Agentic systems magnify these issues because they operate across silos, demanding integration with electronic lab notebooks, LIMS, automation hardware, cloud compute, and regulatory systems that were never designed to talk to each other.
Governance is another open wound. BCG's survey found that 35% of companies are already exploring agentic AI, with another 44% planning near-term deployment, yet governance frameworks and operating models lag the technical rollout by months or years. Who owns the output when an agent autonomously designs a protocol? How do you audit an agentic workflow for regulatory submission? What happens when an agent makes a mistake that cascades through downstream experiments, contaminating weeks of work?
These aren't hypothetical concerns. Nature Reviews Chemistry published a perspective in 2025 calling for rigorous safety frameworks in autonomous experimentation, emphasizing incident prevention, interlocks, and human-in-the-loop controls as agents transition from in-silico simulations to wet-lab execution where real reagents get consumed and real cells die if something goes wrong.
Reproducibility and transparency matter, too—perhaps more than efficiency gains. The EMA and FDA guidance converge on trustworthiness, bias mitigation, and validation. In practice, that means agentic systems need to generate transparent runbooks, maintain provenance for every decision, and version artifacts in ways that satisfy quality systems designed for human experimenters. The research community is responding: benchmarks like LAB-Bench, proposed in 2024, aim to standardize evaluation of language models and agents on biology-specific tasks beyond simple textbook Q&A—literature navigation, protocol generation, database queries. The field is maturing. But it's still early enough that standards are being written in real time, sometimes by the same people trying to commercialize the technology.
Access could become a constraint, maybe a serious one. AlphaFold 3's initial release drew sharp criticism for restricted server access and limited code availability, frustrating researchers who wanted to build on the model rather than just use it. Proprietary tools and data lakes—especially in a field where scale confers compounding advantages—risk creating haves and have-nots. Open models like ESM3 and frameworks like Autoprotocol offer partial counterweights, but the ecosystem remains fragmented enough that interoperability is more aspiration than reality.
The policy environment isn't helping. The U.S. shifted from a broad AI safety executive order to a more laissez-faire stance in early 2025, though domain regulators like the FDA keep issuing guidance documents that don't always align with White House rhetoric. The EU's AI Act includes a research exemption for tools "developed and put into service for the sole purpose of scientific research," but high-risk classification still applies in diagnostics and clinical contexts. Biosecurity adds yet another layer: as model capabilities advance, private labs have quietly tightened usage policies, and international frameworks like the Bletchley Declaration flag dual-use concerns that no one has convincingly resolved. Founders and R&D directors navigating this landscape might be wise to anchor to sector-specific guidance and internal quality systems rather than waiting for top-down clarity that may never crystallize.
What Comes Next
Deloitte's 2025 life sciences outlook found that 60% of executives plan to increase generative AI investments this year, with near-term value expected in R&D, manufacturing, supply chain, and commercial operations. BCG predicts that by 2028—not some distant future, but three years from now—agentic AI will drive roughly 30% of AI value for leading organizations, requiring restructuring of operations and governance to give these systems more decision-making latitude. That timeline should focus minds.
So what does this mean for founders, investors, R&D leaders trying to place informed bets? The opportunity is measurable: a $110 billion value pool in pharma alone, a lab automation market approaching $12 billion, federal infrastructure investments signaling long-term commitment. The technology is proven enough to move beyond demos, but immature enough that the winners aren't locked in yet. The startups building agent orchestration layers, the platforms embedding agents into existing workflows, the infrastructure plays connecting models to hardware—these are the categories worth watching closely.
But the path from pilot to production won't be smooth, and it probably won't be fast. Integration complexity, regulatory uncertainty, the need for entirely new operating models—these will separate experiments from enduring businesses. Agentic biology isn't replacing scientists, despite what some breathless pitch decks suggest. It's changing what scientists do. The labs that figure out how to genuinely collaborate with autonomous systems—how to govern them, trust them, scale them without breaking things—will compound advantages that compound quickly.
The rest will keep pipetting by hand, wondering what happened.
