A robot at the University of Toronto doesn't sleep. It doesn't take coffee breaks or submit grant applications. What it does—and this is the part that unnerves some scientists—is think.
Well, something like thinking. The machine, housed in the university's Acceleration Consortium, selects its own experiments. It analyzes the results, adjusts its hypotheses, and begins again. No human standing over its shoulder. In two months last year, it identified 21 promising organic semiconductor candidates, work that would have consumed years of a researcher's career. The kicker? Materials that once demanded two decades and $100 million in development costs are now being discovered in roughly a year for about $1 million.
This is what people in the field have started calling "agentic biology," and it represents something more profound than faster pipelines or smarter software. These aren't systems that crunch data on command. They make decisions. They plan protocols, interrogate databases, write executable code, command robot arms, and close the entire experiment-design-measure-learn loop—often without pausing for human approval at every turn.
According to a 2025 review of self-driving laboratories, the technology accelerates drug discovery by a median factor of six. Six times faster. That's not incremental. That's a fundamentally different way of doing science, and it's beginning to reshape the economics of biotech R&D in ways that seemed almost fanciful as recently as 2023.
The Money Follows the Machines
The synthetic biology market sits somewhere between $16 billion and $24 billion this year, depending on whose projections you trust. By 2034, forecasters expect that figure to balloon to $80 billion on the conservative end, perhaps $192 billion on the bullish side—compound annual growth rates approaching 17% to 29%. More than 5,400 investors are backing over 1,300 synbio companies through north of 5,600 funding rounds. North America leads, but this is very much a global race.
What's shifted isn't merely the capital. It's the underlying infrastructure. McKinsey and QuantumBlack estimate that AI-accelerated R&D could unlock between $360 billion and $560 billion in annual economic value across sectors, potentially doubling innovation rates in industries where intellectual property matters most—pharma, advanced materials, chemicals. The OECD's 2025 assessment singles out the convergence of synthetic biology, artificial intelligence, and automation as a defining force of the decade ahead. One that, in the OECD's careful phrasing, demands "anticipatory governance."
"Self-driving lab" has entered the vernacular, though the term can be misleading. These are autonomous systems using machine learning to choose experiments and robotics to run them. That median six-times acceleration comes from a 2025 benchmarking study analyzing published results across materials science and chemistry. In some cases, enhancement factors spike to 10 or even 20 experiments per optimization dimension. Not marginal gains—a wholesale reimagining of laboratory tempo.
But the self-driving labs are only half the story. The real transformation is happening at the software layer, where large language models equipped with tools are becoming something closer to lab managers.
The Software Gets Fluent
LLM-driven agents can reason, retrieve information from vast databases, write code that actually runs, generate experimental protocols, and interface directly with electronic lab notebooks, laboratory information management systems, and robotic hardware. They're plugging into the same software infrastructure that human scientists use, only with a fluency that borders on unsettling.
Last October, Anthropic launched Claude for Life Sciences, designed to integrate with platforms like Benchling, PubMed, and Synapse. Eric Kauderer-Abrams, Anthropic's head of life sciences, told TechBrew the goal is for "a meaningful percentage of life sciences work" to eventually run on Claude—everything from hypothesis generation to regulatory documentation. "We aim to eventually close the loop so Claude assists with execution," he said. Not just planning. Execution.
That same month, Benchling announced a partnership with Anthropic to expose its scientific data to Claude through what's called the Model Context Protocol—a standard Anthropic introduced to let AI models connect with external systems while maintaining governance and audit trails. Sanofi, which has 1,500 scientists using Benchling, is building its entire AI strategy on this data backbone. Ashu Singhal, Benchling's co-founder and president, framed it bluntly: "AI in R&D only works through an ecosystem... access, governance, interoperability first."
In January 2026—just weeks ago—Owkin rolled out what it's calling "agentic infrastructure for biology" and formalized its relationship with Anthropic. Around the same time, HighRes and Opentrons announced the first "AI agent-to-agent" lab automation workflow, connecting HighRes's Cellario orchestration software and FlexPod hardware to Opentrons Flex robots and OpentronsAI. A demo is scheduled for SLAS 2026. It's no longer just AI controlling robots. It's AI systems negotiating with one another to execute complex, multi-step workflows.
At an even more fundamental level, researchers are building operating systems purpose-built for autonomous labs. UniLabOS, detailed in a December 2025 arXiv paper, envisions an AI-native OS with typed resources, transactional safeguards, and multi-node coordination across distributed edge-cloud environments. Infrastructure designed from the ground up for machines that conduct experiments.
The scale of data generation is staggering. Ginkgo Bioworks, with its foundry-scale operations, has tested 5 million enzyme designs and performs 100 million multiplex edits per year—volumes no academic lab could dream of matching. That data feeds AI models trained on datasets orders of magnitude larger than what was available just a few years ago.
Google DeepMind's AlphaFold 3, released in May 2024, doesn't just predict protein structures anymore. It models their interactions with ligands, nucleic acids, ions—the molecular choreography that determines whether a drug candidate works. Isomorphic Labs, DeepMind's commercial spinout, has inked partnerships with Eli Lilly and Novartis. In 2024, EvolutionaryScale launched with a $142 million seed round and introduced ESM3, a generative biology model that designed a novel fluorescent protein from scratch. Profluent debuted OpenCRISPR-1, an AI-designed gene editor released as open source.
This isn't speculative work anymore. Generate:Biomedicines announced in December 2025 that GB-0895, an AI-engineered antibody for severe asthma, is advancing to Phase 3 trials. Absci dosed its first AI-designed, zero-shot antibody—ABS-101—in humans last year. These molecules weren't discovered by human intuition or serendipity. They were computed.
From Academic Project to Commercial Infrastructure
Phylo might be the clearest case study in how academic research becomes commercial infrastructure, and fast. Founded by Stanford researchers Kexin Huang and Yuanhao Qu, Phylo raised a $13.5 million seed round in early 2025, co-led by Andreessen Horowitz and Menlo Ventures' Anthology Fund, with participation from—tellingly—Anthropic itself.
Their open-source Biomni agent, introduced in 2025, supports what they call "action discovery" across 25 biomedical domains. It's been adopted widely in academia and industry. Phylo's commercial product, Biomni Lab, borrows a term from software engineering: it's an "Integrated Biology Environment," akin to an IDE for code, but for biological experiments. The system connects AI agents to over 300 tools and claims performance improvements of 20% or more compared to other agent systems (though that figure comes from company materials, not independent benchmarks).
The scientific advisory board includes Carolyn Bertozzi, Feng Zhang, and Fabian Theis—names that carry weight. Jorge Conde from a16z cited "an uncommon level of real-world adoption from academia" as a key driver behind the investment. Translation: this wasn't just promising on paper. Scientists were already using it.
ChemCrow, published in Nature Machine Intelligence in 2024, demonstrated that an LLM agent equipped with 18 chemistry tools could autonomously plan and execute chemical syntheses and discovery tasks. Proof of concept that tool-augmented agents can handle physical workflows, not just digital ones.
Then there's the Acceleration Consortium at the University of Toronto, led by Alán Aspuru-Guzik, which received a $200 million grant from Canada's federal government in 2024 to advance self-driving labs. Aspuru-Guzik's framing is bracingly direct: materials development that once took "20 years and $100 million" can now be done in "around one year and $1 million." The consortium has partnerships with BASF and is building a global SDL network. Those 21 organic semiconductor candidates discovered in two months? A timeline that would have been laughable a decade ago.
The infrastructure layer is equally telling. Emerald Cloud Lab offers fully software-controlled remote wet labs with data provenance meeting ALCOA+ standards—attributable, legible, contemporaneous, original, accurate, plus complete, consistent, enduring, available. In other words, the kind of rigor regulators demand.
Culture Biosciences, which closed a Series C in December 2025, launched the Stratyx 250, a cloud-integrated mobile bioreactor paired with Google Cloud's Gemini AI for real-time analytics. Scientists control fermentation runs from anywhere. The AI monitors, adjusts, flags anomalies. The scientist checks in periodically, but the system doesn't wait for permission to tweak parameters.
NVIDIA partnered with Novo Nordisk in 2025 to develop agentic AI workflows using NVIDIA's BioNeMo models for single-cell modeling and drug discovery. The infrastructure vendors are positioning themselves not as tool suppliers anymore. They're building orchestration platforms—systems that let disparate agents and robots communicate, like an air traffic control system for laboratory automation.
The Friction Points
The trajectory is steep. But it's not frictionless, and the problems are starting to reveal themselves.
BioAgent Bench, a benchmark introduced in January 2026, tests frontier LLMs on end-to-end bioinformatics pipelines. The models can complete tasks, yes. But they show brittleness under perturbations. Minor changes—a slightly different file format, an unexpected error message—break workflows entirely. That's a problem in regulated environments where robustness isn't a nice-to-have. It's table stakes.
Data governance is messy. Benchmarks reveal that closed models may be fundamentally unsuitable for research with strict privacy requirements, pushing some institutions toward governed, open-weight alternatives. Non-deterministic agent behavior complicates validation. You can't just re-run the experiment and expect identical results when the AI makes probabilistic decisions. Runtime analytics beyond black-box performance scores are still emerging. And reproducibility—the bedrock of scientific credibility—gets murky when agents make choices that aren't fully transparent.
Then there's biosecurity, which is perhaps the most consequential friction point. The OECD, the U.S. Office of Science and Technology Policy, and the UK government have all flagged AI's dual-use potential. In 2024, OSTP released a unified policy for oversight of dual-use research of concern and pathogens with pandemic potential, effective May 2025. A 2025 executive order mandates enforcement mechanisms for non-federally funded risky research and comprehensive nucleic-acid synthesis screening.
Studies have shown that LLMs can assist in designing harmful molecules and proteins. Not hypothetically. They've done it in controlled tests. That capability demands multi-layered mitigations—technical, regulatory, and cultural.
A security incident in December 2025 involving Anthropic's Model Context Protocol underscored the composite-system risks. When you interconnect multiple AI systems, databases, and robots, vulnerabilities don't just add up. They compound. One weak link can unravel the whole chain.
What Comes Next
Still, the momentum is hard to ignore.
More MCP-based integrations are linking electronic lab notebooks, LIMS, knowledge graphs, and automation systems. The Linux Foundation's Agentic AI Foundation—with contributions from Anthropic and OpenAI—is working to standardize agent protocols, MCP among them. The goal is open, neutral infrastructure that prevents vendor lock-in and encourages interoperability. Whether that vision materializes is anyone's guess, but the effort is real.
Wider adoption of self-driving labs in pharma and advanced materials is generating real-world acceleration data. More AI-engineered biologics are entering mid- and late-stage trials, which means we'll soon know whether these molecules are as good as their computational promise suggests. Regulatory frameworks are tightening around high-risk biological research, including how agent platforms broker lab execution and who can order what DNA sequences under what oversight.
Eric Kauderer-Abrams's vision of Claude eventually assisting with execution doesn't sound far-fetched anymore. Agent-to-agent pipelines like the HighRes-Opentrons collaboration suggest that closing the loop—AI deciding and doing—is a near-term reality, not a distant ambition.
The self-driving lab isn't coming. It's here. Running overnight shifts, adjusting parameters in real time, discovering molecules that human intuition might never have considered. Perhaps more molecules than any human ever could.
The question isn't whether AI agents will transform biotech R&D. It's how quickly the industry can build the governance, standards, and trust required to deploy them safely at scale. And whether, when those systems inevitably fail in unexpected ways, we'll be ready for what breaks.
