AI is moving off the screen and into the lab.

Ginkgo opened its autonomous robotic infrastructure to anyone with a browser - the same fleet that ran 36,000 experiments with OpenAI's GPT-5, now available starting at $39.

Stanford has shipped three different agent systems: PantheonOS connects AI agents to 22 biology foundation models and lets them evolve their own algorithms; Eubiota screened nearly 2,000 genes for stress tolerance, designed a gut bacteria therapy, an antibiotic cocktail, and anti-inflammatory molecules - all validated experimentally in the lab; and LabOS put AI behind smart glasses to catch scientists' mistakes in real time, training junior researchers to expert level in a week.

Meanwhile, Arc Institute's Evo 2 landed in Nature with experimental proof that its AI-designed DNA sequences work in living cells - including what Arc calls the first AI-designed organisms.

Every story in this issue involves AI designs that were tested in actual labs, in actual cells, in actual mice - and held up.

Let’s dive in.

Ginkgo Bioworks launched Cloud Lab - a web interface that gives outside researchers direct access to the company's autonomous lab infrastructure in Boston. Scientists submit a protocol in plain language, and an AI agent called EstiMate assesses whether Ginkgo's robotic fleet can run it, then returns a price quote. The system runs on Ginkgo's Reconfigurable Automation Carts (RACs) - modular units combining robotic arms and maglev sample transport - with access to over 70 instruments covering sample prep, liquid handling, and analytical readouts.

Cloud Lab runs on the same RAC-based autonomous infrastructure in Boston that ran 36,000 experiments with OpenAI's GPT-5 - the collaboration we covered in Issue 1 - now opened up to external researchers. If the The OpenAI news was a showcase, Cloud Lab is the commercial pitch: anyone can now submit work to the same robotic fleet.

The timing is interesting. Cloud Lab arrived days after Ginkgo reported Q4 earnings showing cell engineering revenue down 26% year-over-year to $26 million, with full-year 2025 revenue at $133 million versus $174 million in 2024. Cash burn improved - down 55% year-over-year to $171 million - but the company is guiding on cash burn rather than revenue for 2026, signaling it's prioritizing investment over near-term growth. Ginkgo is also divesting its biosecurity business to concentrate capital on autonomous labs, and announced it will decommission traditional lab benches in Boston, migrating all R&D services onto its Nebula autonomous lab.

CEO Jason Kelly framed Cloud Lab as starting small - experiments starting at $39 - to let scientists try autonomous lab work before larger commitments.

Why it matters: Ginkgo's revenue has declined for three straight years and the stock is down more than 95% from its 2021 peak. Cloud Lab is the bet that turns years of autonomous lab investment into a platform business - lab-work-as-a-service starting at $39 per experiment. Whether outside researchers actually show up will say a lot about whether the autonomous lab model can scale beyond partner collaborations like the OpenAI project.

Did you know? Ginkgo's Nebula lab in Boston currently has over 50 RAC units and plans to expand to 100 in the first half of 2026. You can submit a protocol and get a feasibility report today at cloud.ginkgo.bio.

Xiaojie Qiu's lab at Stanford - the group behind single-cell analysis tools Monocle and Dynamo - has released PantheonOS, a multi-agent framework that connects LLM-powered agents to more than 22 foundation models for biology. A preprint posted to bioRxiv describes a distributed system where specialized agents collaborate across machines to run end-to-end computational biology workflows. The team calls it "vibe analysis" - vibe coding for science.

The key trick is a sort of self-improvement. A module called Pantheon-Evolve takes existing bioinformatics tools and evolves better versions of their code using a genetic algorithm - breeding successive generations until performance improves. Applied to two standard data-cleaning methods, the evolved variants outperformed the originals.

In case studies, PantheonOS drove actual biological findings: it found how signaling proteins are unevenly distributed in early mouse embryos, helping explain how embryos establish their body axis. It mapped molecular programs underlying heart disease in fetal heart data, and adaptively selected virtual cell models to predict gene regulatory networks and perturbation effects in the developing heart. These findings have not yet been peer-reviewed.

The platform runs entirely offline with local LLMs - a privacy-by-design choice for labs working with sensitive patient data.

Why it matters: This is the third Stanford multi-agent biology system BAIO has covered in five issues - Virtual Lab, Virtual Biotech, and now PantheonOS. Where James Zou's systems mirror organizational structures, Qiu's builds an operating system that connects agents to foundation models and lets them evolve their own algorithms.

Did you know? PantheonOS supports Python, R, Julia, and LaTeX and can be installed with pip install pantheon-agents. Qiu received an NIH High-Risk High-Reward grant in 2024 specifically to build virtual embryo models - the same kind of data PantheonOS is now analyzing.

James Zou's lab at Stanford (the same Zou mentioned above) - in collaboration with microbiome researcher Justin Sonnenburg - has released Eubiota, an open-source AI agent framework for gut microbiome discovery. A preprint posted to bioRxiv describes a system of four specialized agents that coordinate through shared memory to run multi-step scientific investigations.

The standout result: Eubiota drove four discoveries, each validated experimentally: it screened 1,945 genes across 10,800 papers in hours on two GPUs to find which ones help gut bacteria survive inflammatory stress; designed a four-species microbial therapy that reduced colitis in mice; engineered an antibiotic cocktail that kills pathogens while sparing beneficial gut bacteria; and discovered novel anti-inflammatory metabolites from a large human dietary study. These findings have not yet been peer-reviewed.

Why it matters: Many AI agent papers show agents answering questions. Eubiota closes the loop: the agent designs the experiment, humans run it in the lab, and the results hold up across four different biological challenges. That the system runs on an open-weight 8B model and still outperforms GPT-5.1 in blinded expert evaluation suggests architecture matters more than model size.

Did you know? Eubiota's interactive platform is free at app.eubiota.ai, and the full system is open-source on GitHub. For labs working with sensitive patient data, it supports local deployment behind institutional firewalls - all agents can run on open-weight models with no external API calls required.

Le Cong's lab at Stanford and Mengdi Wang's at Princeton have built LabOS - an AI system that watches scientists work through smart glasses and catches mistakes in real time. Powered by a custom vision-language model trained on 200+ expert-annotated lab video sessions, the system streams video from AR glasses, compares what it sees against the written protocol, and guides the wearer step by step. The largest model achieves over 90% error detection accuracy.

In a protein expression experiment reported by Scientific American, junior scientists with one week of LabOS training produced results their PI couldn't distinguish from expert work. The system works by recording expert practice through the glasses, then coaching novices through the same procedure - flagging sterile technique lapses, wrong reagent incubation times, and skipped steps as they happen.

Why it matters: LabOS puts AI into the physical lab, watching what scientists do with their hands. If reproducibility failures are largely human protocol errors - a 2016 Nature survey found over 70% of scientists couldn't reproduce others' experiments - real-time AI guidance could address the problem at its source.

Did you know? The team has already expanded into surgery with MedOS, announced in February. Using the same smart glasses and AI architecture, early results show nurses improving from 49% to 77% performance with AI assistance. MedOS is being showcased at NVIDIA GTC this month.

Arc Institute's Evo 2 has been published in Nature. Arc trained two model versions - 7B and 40B parameters - both with a 1 million token context window, on 9.3 trillion nucleotides from over 128,000 genomes spanning all three domains of life.

The model, developed with NVIDIA and collaborators at Stanford, UC Berkeley, and UCSF, predicts functional impacts of genetic variation - including noncoding pathogenic mutations and clinically significant BRCA1 variants - without task-specific fine-tuning.

What's changed since the preprint a year ago: the designs are working in the lab. Chromatin accessibility patterns generated by Evo 2 closely matched experimental measurements when synthesized and inserted into mouse stem cells. And 16 of 285 AI-designed bacteriophage genomes successfully killed target bacteria without affecting unrelated strains - what Arc calls the first experimentally validated AI-designed organisms.

The model is fully open-source - weights, training code, inference code, and the OpenGenome2 training dataset - and integrated into NVIDIA's BioNeMo framework. Since the preprint, it's been downloaded over 88,000 times on GitHub and received more than 8 million API requests on Hugging Face.

Why it matters: Evo 2 going from preprint to Nature is validation for the biological foundation model field. Patrick Hsu, Arc's co-founder, says the next step is designing entire gene and protein collections without explicit rules - moving toward what he calls a "read-write-think loop" for biology.

Did you know? Arc is hiring.

If you use Claude Code, Cursor, or Codex for research, K-Dense has open-sourced a collection of 170+ scientific skills that turn your AI coding agent into a lab assistant wired to real databases. The skills cover bioinformatics (Scanpy, BioPython, scvi-tools), cheminformatics (RDKit, ChEMBL), clinical research, and more. Query ChEMBL for drug candidates, run single-cell RNA-seq workflows, or search PubMed - all from your terminal.

☑️ Science asks how we'll know when AI is smart enough to do science.

☑️ OMICmAge extends GrimAge with a new multi-omic aging clock from ~31,000 participants, published in Nature Aging.

☑️ Asimov Press on the legibility problem: what happens when AI makes scientific discoveries humans can't understand?