The researchers at Google didn't use the word "replace" when they unveiled their AI co-scientist last February. They used "collaborate." The system—built on Gemini, naturally—generates hypotheses and research plans on its own, then waits for a human to sign off. It's careful positioning, the kind that acknowledges what everyone in the room is thinking but nobody wants to say out loud.
A YC-backed startup called Synthetic Sciences has fewer qualms. When it presents at Demo Day later this month, the pitch will be more audacious: agents that orchestrate other AI co-scientists to handle every stage of research, from literature review through experiment execution to manuscript drafting. The humans? Optional for large stretches of the process.
Whether this is premature hype or the opening salvo of something genuinely transformative remains unclear. What's certain is that the race is underway, and the money pouring in suggests investors believe the finish line is closer than skeptics think.
Follow the Money, If Not the Hype
The numbers make their own argument. AI in drug discovery was pegged at roughly $1.7 billion in 2024; market researchers project it will hit $8.5 billion by 2030—a compound annual growth rate around 31%. Lab automation sits somewhere between $8.27 billion and $9.05 billion today, trending toward $14.66 billion by decade's end. McKinsey, in a January 2024 estimate, suggested generative AI could deliver $60 billion to $110 billion annually to pharma and medical products alone, primarily by accelerating discovery timelines and streamlining clinical documentation.
Those figures tell one story. Deloitte's April 2024 pharma innovation report tells another. Yes, Big Pharma's R&D return on investment rebounded to 4.1% from a dismal 1.2% in 2022. But the report landed a blunt assessment: "AI is yet to be a game-changer" at scale. The gap between what venture capital believes and what lab benches deliver remains uncomfortably wide.
When Assistants Become Autonomous
The tools themselves are evolving past their original design. Early AI applications in biopharma were tightly scoped—predict a protein structure, generate a candidate molecule, search the literature. AlphaFold 3, published in Nature last May, represented a leap: a unified diffusion model capable of predicting protein-nucleic acid-ligand complexes in concert. DeepMind's GNoME system, announced in November 2023, predicted 2.2 million new crystal structures, though debates over novelty and real-world validation continue.
Now the industry is chasing what insiders call "agentic workflows"—multi-stage systems that don't just assist but orchestrate. They integrate retrieval, code execution, experiment control, document drafting. Google's AI co-scientist embodies the shift. The system deploys six specialized agent roles (Generation, Reflection, Ranking, Evolution, Proximity, Meta-review) to produce research plans that humans then vet. Google has been careful, almost defensive, in positioning it as collaborative rather than autonomous. Its February and December 2025 blog posts emphasized rigorous evaluation and factuality tools, perhaps anticipating the backlash that came anyway.
Synthetic Sciences, founded in 2025 and based in Mountain View, takes a different approach. The startup markets five specialist agents: Research (a seven-stage pipeline), Flywheel (for LLM training), Write (LaTeX with citations and figures), Code (general coding), and Plan (read-only exploration). The platform orchestrates external compute services—Modal, Lambda, Tensorpool, Groq, Together AI, Fireworks—and integrates with HuggingFace, Weights & Biases, and others.
Pricing is unusually transparent for the space: $50 a month for Plus, $250 for Pro, custom quotes for Enterprise. That buys access to three frontier models, GPU orchestration, and a cloud runtime designed for long-running agents. The company claims it builds a knowledge graph of each user's work, with agents that "never stop." Founder details haven't been made public yet; the company is part of YC's Winter 2026 batch and hasn't given interviews ahead of Demo Day.
Compared to Google's internal research program or established platforms like Benchling—which embeds AI assistants into electronic lab notebooks and LIMS—Synthetic Sciences appears to target researchers who want pre-packaged agent orchestration without managing infrastructure themselves. Whether that market is large enough to justify venture-scale returns is an open question.
The Labs That Run Themselves
The agents aren't only planning research anymore. They're starting to execute it.
A growing number of "self-driving labs" couple AI with robotics to run closed-loop experiments. The A-Lab at Lawrence Berkeley National Laboratory, described in Nature in November 2023 (with a correction issued in January 2026), synthesized 36 of 57 target inorganic materials in 17 days, integrating ab-initio calculations, text-mined synthesis heuristics, active learning, and robotic powder handling. Argonne National Laboratory's "Polybot," announced in February 2025, automates polymer film discovery for electronics. Researchers at North Carolina State reported last July that a self-driving lab using dynamic flow conditions generated ten times more data than steady-state approaches, meaningfully accelerating candidate identification.
Axios framed self-driving labs as a "strategic national capability" in an August 2024 piece, noting a proposed FASST initiative that's still pending in Congress. The Department of Energy and NSF are backing testbeds. Emerald Cloud Lab offers a commercial remote-controlled life sciences facility, pitching ALCOA+ data integrity and academic time savings to universities struggling with equipment costs.
Benchmarking efforts are gaining momentum too. HeurekaBench, released in January 2026 as an open framework for evaluating AI co-scientists in single-cell biology, reported that a critic module improved open-source agent quality by up to 22%. Academic workshops are proliferating: an ICLR 2025 session on agentic AI for science, and the Agents4Science 2025 conference organized by Stanford's James Zou, which explores AI as authors and reviewers. The latter raised eyebrows—the idea of AI reviewing AI-generated science makes some researchers queasy.
A Crowded, Chaotic Field
The competitive landscape defies easy categorization. On the drug discovery side, Isomorphic Labs—DeepMind's pharma spinout—inked multi-target deals with Eli Lilly and Novartis in 2024 with headline milestones approaching $3 billion. The company reportedly raised $600 million led by Thrive in 2025, though neither confirmed exact figures. Insilico Medicine advanced the first fully generative-AI drug into Phase II (Rentosertib for idiopathic pulmonary fibrosis), posting positive Phase IIa data in November 2024 and securing a USAN name this past March. Exscientia is expanding its automated "Design-Make-Test-Learn" loop on AWS, with trials planned for 2025. Recursion completed BioHive-2, a 63-unit DGX H100 supercomputer it calls the largest wholly owned biopharma AI system, backed by a $50 million strategic investment from NVIDIA in 2023.
Software platforms are embedding agentic features too. BenchSci announced a three-year global license with Sanofi in October 2025 for its ASCEND neurosymbolic AI platform. Causaly unveiled Agentic Research in the fall, positioning domain-specific agents with Bio Graph and Scientific IR for traceable, multi-step R&D workflows. Benchling rolled out AI Compose, Data Entry, Ask Mode, and Deep Research at its Benchtalk 2025 conference, integrating NVIDIA's BioNeMo and OpenFold2 via NIM and adding Python and R execution in-platform. At the conference, executives emphasized that "AI becomes truly powerful when embedded directly in scientists' workflows"—a not-so-subtle jab at standalone agent platforms.
FutureHouse, a non-profit, and its commercial spinout Edison Scientific (launched in November 2025) claim an integrated AI Scientist capable of literature search, data analysis, hypothesis generation, and experimental planning. Opentrons continues to democratize lab automation with accessible liquid handlers like the OT-2 and Flex. The Materials Project, a DOE-backed database, released datasets like MatSyn25 (synthesis steps for 2D materials) in January 2026 and LeMat-Traj (large trajectory data) last August, fueling open research.
It's a lot of companies chasing what might be the same problem. Or different slices of it. Nobody's quite sure yet.
The Skeptics Have Receipts
Not everyone is convinced the technology is remotely ready. TechCrunch gathered academic critiques in March 2025 under the headline "Experts Don't Think AI Is Ready to Be a Co-Scientist," citing concerns about hallucinations, junk-science proliferation, and the persistent gap between computational predictions and physical-world constraints. Google's own blog posts acknowledge limitations and emphasize the need for subject-matter expert review—hardly a ringing endorsement. Deloitte's report was more direct: AI hasn't delivered the promised game-change at scale, full stop.
Regulatory pressures are mounting in ways that could reshape the field. The NIH prohibits AI tools in peer review and, as of a July 2025 clarification, will not consider applications "substantially developed by AI" (effective for September 25, 2025 receipt dates). Leading journals disallow AI as an author and require disclosure for AI-assisted text or figures. The EU AI Act, finalized in 2024, imposes systemic-risk obligations on general-purpose AI providers, with compliance implications for scientific agent platforms operating in Europe.
U.S. policy is tightening too. A May 2024 OSTP update expanded oversight of dual-use research and pathogens with enhanced pandemic potential. HHS 2023 guidance on synthetic nucleic acid screening now requires federal funders to mandate adherence starting April 26, 2025—a deadline that's caught some labs off guard.
Access debates are intensifying. AlphaFold 3's restricted availability versus AlphaFold 2's open model drew pushback from researchers frustrated by licensing constraints, as Le Monde reported last May. The question of who controls foundational scientific AI—and under what terms—cuts to the heart of how discovery will scale, or whether it will at all.
What the Smart Founders Are Watching
The landscape is consolidating around a few technical and business patterns. Domain-specific knowledge graphs and neurosymbolic approaches (BenchSci's BEKG, Causaly's Bio Graph) are gaining traction for improving traceability and regulatory acceptability. Integration of foundation models like AlphaFold 3 and NVIDIA's BioNeMo directly into R&D platforms is accelerating. Data provenance by design—ALCOA+ compliance, audit trails, transparent agent decisions—is shifting from nice-to-have to procurement requirement, particularly for pharma buyers who've been burned before.
Expect more sector-specific benchmarks following HeurekaBench's lead, pressure for reproducible case studies rather than cherry-picked demos, and an emphasis on agent competitions. Pharma-scale licenses like Sanofi's deal with BenchSci point toward enterprise consolidation, where a handful of platforms capture the bulk of revenue while smaller players scramble for niches. Self-driving labs will expand via DOE and NSF testbeds and private cloud-lab integrations, though chemistry and hardware remain gating factors—software can only move as fast as pipettes.
Clinical proof points will matter most. Insilico's Phase II data is a start, but the field needs multiple AI-discovered molecules reaching late-stage trials before the ROI narrative becomes undeniable. Until then, differentiation will hinge on workflow integration, compliance tooling, and the ability to demonstrate reproducible, auditable research at scale. Not flashy demos.
The Messy Reality Ahead
The race to build AI co-scientists isn't over. It's only just beginning to encounter the messy reality of wet labs, regulatory frameworks, and skeptical scientists who've watched too many overhyped technologies fail to deliver.
Whether Synthetic Sciences and its peers succeed depends less on their agent architectures than on whether they can bridge the chasm between computational promise and experimental proof. For biotech founders paying attention—and by now, most are—the question isn't whether these tools will matter. It's how to separate signal from noise before the hype cycle inevitably turns, as hype cycles always do, and the field is left to reckon with what actually works.
The humans in the lab coats aren't going anywhere just yet. But they might, eventually, need fewer colleagues.
