AI Agents Are Doing Science — Where They Actually Help, and Where They Don't

TL;DR

AI agents can now read the literature, generate and rank hypotheses, write and run code, and even drive robots — and in the last two years several have produced results validated in real wet labs. But read past the abstracts and a consistent pattern appears: agents win the cheap, upstream half of science — conjecture, synthesis, design-space search — while the bottleneck has moved downstream to validation, which is still physical, slow, and human. I build scientific agents for a living, so here’s the honest map: where they genuinely help, where the headline numbers quietly soften, and why judgment — not ideas — is the hard part.

“What can you do for me?”

Few people are further out on the frontier of agentic science than James Zou. A Stanford professor of biomedical data science, Zou doesn’t just theorize about AI scientists — his lab builds the most concrete ones going: the Virtual Lab, where a GPT-4 “principal investigator” convenes a team of specialist agents and a critic to run real research [3], and Paper2Agent, which turns a published paper’s code into a callable, reliable tool [14]. When someone building at that level tells you where the field is headed, it’s worth reading closely.

Which is why a recent Stanford Medicine piece, where Zou lays out the shift, stuck with me. He frames it cleanly [1]: AI is moving from “what can you tell me?” to “what can you do for me?” — systems that reason, pursue a goal, use tools, and act with minimal supervision. And his caveat is the honest part — agents “still make mistakes,” and physical lab testing “remains essential” [1].

That caveat is the story. Let me unpack it from the builder’s side.

Where agents actually help: the cheap half of science

Every substantiated result I can find lands in the same band — the front half of the research funnel:

• Literature synthesis. FutureHouse’s PaperQA2 beats PhD scientists on answer precision (85.2% vs 73.8%) on a hard literature-QA benchmark (LitQA2) [4] — precision, note, not overall accuracy, since it can abstain when the evidence is thin. Honestly, this is the single most useful thing agents do for me today.
• Hypothesis generation and ranking. Google’s AI Co-Scientist proposed hypotheses that held up in the wet lab — an AML drug-repurposing hit (KIRA6, IC50 13 nM in one line) and anti-fibrotic targets validated in human liver organoids [2]. (A third showcased case only recapitulated an already-known but unpublished result, so it carries less evidentiary weight.)
• In-silico design. Zou’s Virtual Lab agents assembled a working ESM → AlphaFold-Multimer → Rosetta pipeline and designed 92 nanobody variants [3].
• Reliable method reuse. Paper2Agent turns a paper + its code into a locked toolset — 22 callable tools from AlphaGenome in ~3 hours, no human in the loop [14]. Turning a method into something you can call without re-deriving it is a quiet, high-value win.

DeepMind’s Conjecture Machines essay names the pattern: agents make hypotheses “abundant and cheap” [8]. That is the honest answer to “where do they help” — the cheap, upstream stages.

Read the supplement

Here’s the discipline the hype skips, and the one I’d underline hardest: almost every headline number softens when you read the primary source.

• A-Lab’s famous “41 new materials in 17 days” became 36 of 57 in a 2026 Nature correction [6] — and outside chemists argue the count of genuinely-novel compounds is closer to zero, because the robot’s automated X-ray characterization mis-called what it actually made [7].
• The Virtual Lab’s “functional nanobodies” headline is 2 of 92 with meaningful new binding, at moderate affinity [3].
• Robin’s preprint reported a large automated effect for its lead candidate that shrank sharply under human re-analysis; by the final Nature paper the automated (1.89-fold) and human (1.75-fold) estimates had converged [11]. The lesson holds — automated analysis needs a human auditor — but here peer review is what caught it.
• Kosmos reports that 79.4% of its statements are accurate — i.e. ~1 in 5 is wrong [12] — and an independent audit found 1 of 3 of its hypotheses outright false when tested against a null model [13].

None of this is fraud. But the gap between the abstract’s framing and the supplement’s numbers is consistent and large — and if you build on these systems, that gap is your problem to manage.

The bottleneck moved: validation, not ideas

Here’s the part I feel most, building these things: agents are strong at generating and executing, and weak at judging.

• ChemCrow’s most important result is a negative one — GPT-4-as-evaluator couldn’t tell clearly-wrong answers from correct ones; it preferred fluent hallucinations [15]. So much for “LLM-as-judge.”
• The AI Scientist’s “exceeds a top-conference acceptance threshold” is graded by its own automated reviewer — a circular benchmark [9]. An independent audit found 42% of its experiments failed on coding errors alone, and it flagged textbook ideas as novel [10].
• A-Lab could run ~21 experiments a day and still misidentify its products, because automating the measurement and interpretation is the weak link — not moving the matter [6].

This dissolves the “virtual lab vs. self-driving physical lab” debate into a single gap. Virtual agent-labs (Zou’s, Co-Scientist, Robin, SciAgents [17]) scale ideation almost for free but hit a wall at physical reality — humans run every wet-lab step, and that’s where candidate counts collapse (92→2, 15→2). Physical self-driving labs (A-Lab, Coscientist [5], the SAMPLE protein lab) actually move matter and win on throughput and sample-efficiency — SAMPLE found ≥12 °C stability gains while sampling under 2% of the design space [16] — but they break on automated characterization. Both bottlenecks are the same underlying thing: trustworthy, autonomous evaluation of whether a result is real. Today, that loop is still a human.

What this looks like from where I sit

I build scientific agents — for single-cell perturbation prediction, for cardiac drug discovery, for hard math. So “where do agents help” isn’t abstract for me; it’s a set of operating principles I’ve had to learn the hard way. Here’s how I think the human’s job changes.

The human decides what and why; the agent handles how. AI doesn’t need a cancer drug — we do. Think of Waymo: the car drives beautifully, but two jobs stay ours — set the destination, and choose the route. Get those wrong and it arrives, on time, at the wrong place. Once the destination is right, it doesn’t matter whether it took lane 1 or lane 2. Same with code: if it does the job, if/else vs try/except is noise. The machine supplies capability; you supply the goal, the meaning, and the values — it is astonishingly capable and completely amoral.

Your attention is the scarce resource — attention is all we need. The marginal cost of a first attempt — a draft, a literature pass, first-pass code — is collapsing; the cost of proving it correct is not. Picture a magic wand in your pocket: the hard part isn’t the magic, it’s knowing what to ask for. When execution is cheap, the whole bottleneck moves to direction and judgment — the two things you can’t delegate. And in open-ended research there’s no predefined path; you often don’t even know a solution exists. Every step is a decision about which question is worth exploring. That taste — choosing the right problem — is the leverage now, not answering well.

Not trust — verification, with references. This is where the research above gets personal. Every claim an agent makes should trace to a source or an experiment, and I never accept an answer without asking how confident are you, and grounded in what? The deepest failure across all these systems isn’t bad ideas — it’s that agents don’t know when they don’t know (ChemCrow’s own evaluator couldn’t tell wrong from right [15]). So verification can’t be heroic manual checking; it has to be a system you engineer — TDD, gates, adversarial checks, references required — because eyeballing every output doesn’t scale.

Use the right tool for the right job. Not everything is a language-model problem. Open-ended reasoning → the model. Anything that must be exactly right → deterministic code, written test-first (doc → spec → test → code). Facts → retrieval from a grounded knowledge base, not the model’s memory. Most agent failures I see come from using the model where a compiler or a database should have been.

Ground it in an expert-curated, agent-native knowledge base — one that compounds. The other half of “verify, don’t trust” is what you verify against. I build a curated wiki with hybrid search (qmd) so agents reason over vetted knowledge, not the internet’s average. And the point is that it compounds: every verified finding goes back in, so the system gets smarter each run instead of starting cold. That memory is the difference between a clever demo and a discovery engine.

Treat it like engineering, and keep the human on the PR. Agentic work is still software: TDD, small reversible steps, documentation before code. The human reviews the pull request — always. I run a fleet of agents like a manager, not a coder: delegate the execution, keep the direction, the judgment, and the sign-off human.

The output is only as good as you are. The agent is already IQ 500; it will act like IQ 100 if the operator is. Car versus driver — the same F1 car is a trophy or a wreck depending on who’s behind the wheel. So the obligation runs both ways: to get the best from the most powerful thinking machine ever built, I have to keep getting sharper myself. The ceiling on agent output is the human holding the wheel.

One honest caveat. I said syntax doesn’t matter as long as the code works — and for engineering that’s true. But science is different: sometimes the how is the result. A wrong method that produces a right-looking answer is exactly the A-Lab failure — 21 experiments a day, confidently mis-identified [6]. So for building, trust the destination; for discovery, you still have to audit the route.

What’s overhyped — and what isn’t

Overhyped: “autonomous discovery.” In every wet-lab-validated case here, a human curated which agent outputs to test and ran the experiments; “semi-autonomous, expert-in-the-loop” is the accurate label. “AI passed peer review” (AI Scientist-v2) was one coordinated, IRB-approved, workshop-level paper (average score 6.33) that was withdrawn before publication — Sakana itself says none of its papers met a main-conference bar [9].

Not hyped enough: the boring, reliable wins. An agent that reads 150 papers and hands you a ranked, cited shortlist; a method paper turned into a tool you can call; a design-space search that samples 2% and still finds the peak. These don’t make headlines, and they change how I work every week.

What to watch

Whoever closes the verification loop — trustworthy, autonomous evaluation of whether a result is real — unlocks the next step, not whoever ships a flashier ideation agent. Watch three things: better calibration and “epistemic humility” so an agent knows when it doesn’t know; automated characterization you can actually trust (the A-Lab lesson); and null-model auditing built into the agent, not bolted on after. The day an agent reliably says “I’m not sure, and here’s why,” the assistant becomes a colleague.

Sources

1. Stanford Medicine Insights, “Agentic AI in biomedical research” (H. Armitage, interview with James Zou, 2026) — https://med.stanford.edu/news/insights/2026/07/agentic-ai-in-biomedical-research.html
2. Gottweis, Weng, et al., “Towards an AI Co-Scientist” (Google, 2025) — https://arxiv.org/abs/2502.18864
3. Swanson et al., “The Virtual Lab: AI Agents Design New SARS-CoV-2 Nanobodies…” (Zou lab, 2024) — https://doi.org/10.1101/2024.11.11.623004
4. Skarlinski et al., “Language agents achieve superhuman synthesis of scientific knowledge” (PaperQA2, FutureHouse, 2024) — https://arxiv.org/abs/2409.13740
5. Boiko et al., “Autonomous chemical research with large language models” (Coscientist, Nature, 2023) — https://doi.org/10.1038/s41586-023-06792-0
6. Szymanski et al., “An autonomous laboratory for the accelerated synthesis of inorganic materials” (A-Lab, Nature, 2023) — https://doi.org/10.1038/s41586-023-06734-w · Author Correction (2026) https://doi.org/10.1038/s41586-025-09992-y
7. Palgrave & Schoop, A-Lab critique — Nature News https://doi.org/10.1038/d41586-023-03956-w
8. Wallace et al., “Conjecture Machines: AI agents and the new validation bottleneck in science” (Google DeepMind, 2026) — https://deepmind.google/public-policy/conjecture-machines-ai-agents-and-the-new-validation-bottleneck-in-science/
9. Lu et al., “The AI Scientist” (Sakana AI, 2024) — https://arxiv.org/abs/2408.06292 · Yamada et al., “The AI Scientist-v2” (2025) — https://arxiv.org/abs/2504.08066
10. Beel et al., “Evaluating Sakana’s AI Scientist” (independent, 2025) — https://arxiv.org/abs/2502.14297
11. Ghareeb et al., “A multi-agent system for automating scientific discovery” (Robin, FutureHouse) — preprint https://arxiv.org/abs/2505.13400 · Nature (2026) https://doi.org/10.1038/s41586-026-10652-y
12. Mitchener et al., “Kosmos: An AI Scientist for Autonomous Discovery” (Edison Scientific, 2025) — https://arxiv.org/abs/2511.02824
13. Nusrat & Nusrat, “When AI Does Science” — independent KOSMOS evaluation (2025) — https://arxiv.org/abs/2511.13825
14. Miao et al., “Paper2Agent” (Zou lab, 2025) — https://arxiv.org/abs/2509.06917
15. Bran et al., “ChemCrow” (Nature Machine Intelligence, 2024) — https://doi.org/10.1038/s42256-024-00832-8
16. Rapp, Bremer & Romero, “Self-driving laboratories to autonomously navigate the protein fitness landscape” (SAMPLE, Nat. Chem. Eng., 2024) — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10926838/
17. Ghafarollahi & Buehler, “SciAgents” (Advanced Materials, 2025) — https://doi.org/10.1002/adma.202413523

Canonical: https://latentcell.ai/posts/agents-doing-science