AI Is Starting to Grow Hands
Claude no longer stops at reading papers and writing code. Anthropic has built a real biology wet lab where AI directs robots to run experiments, and if AI moves from brain to hands, the production function of science itself could change, not just drug discovery.
However smart AI gets inside a computer, biology's final answer comes from reality
Until now, conversations about AI and drug discovery have mostly stayed inside the computer. Reading papers, predicting protein structures, generating candidates, designing molecules, running simulations. At this stage, AI is already cutting researchers' thinking costs sharply.
But biology has a final step that cannot be skipped. You have to actually try it. Whether a cell responds, a protein binds, or a compound produces the intended effect gets confirmed by a real-world experiment. If AI can generate 100,000 hypotheses a day but a lab can only test 100, the pace of the entire research effort is capped at 100.
That is what makes Anthropic's wet lab interesting. Reuters reported on September 18 that Anthropic has built a wet lab for physical biology experiments in the San Francisco Bay Area. Eric Kauderer-Abrams, the company's head of life sciences, confirmed it. The company is testing whether Claude can direct a robotic unit to run experiments with limited human intervention. At the same time, Anthropic was careful to say this is still "very early stage" and that human oversight and intervention remain essential for safety.
There is an important factual boundary here too. Anthropic told Reuters the wet lab is not a dedicated drug-discovery facility. The company is expanding preclinical drug programs and life-science automation, but calling the whole wet lab a "factory where Claude makes new drugs" would be an overstatement.
Science's bottleneck is shifting from thinking to doing
The closed loop of AI-native science
- Generate a hypothesis
- Design the experiment
- Run it with robots
- Measure and analyze
- Form the next hypothesis
What matters is not the automation rate of each step but how fast the result of one experiment feeds into the next. The shorter this closed loop gets, the more experiments a single researcher may be able to oversee.
Anthropic has already turned this link into products and standards. In June, it released Claude Science, letting researchers handle databases, code and compute environments in one place. In August, it opened the Model Hardware Standard (MHS) for research use, a protocol that lets AI agents safely operate physical equipment like microscopes, liquid handlers and robotic arms.
In one MHS proof-of-concept, Claude coordinated a liquid handler, a robotic arm and a plate reader to run a specific dose-response experiment about 3 times faster. In a separate case, it took roughly 8 hours to set up the equipment integration and orchestration layer, work Anthropic says could take weeks under a conventional vendor approach. These numbers cannot be generalized into a speedup for science as a whole. But it matters that a structure already works in which AI doesn't just write the experimental plan but reads instrument status, issues commands, receives results and decides the next action.
Yesterday's 26% and today's wet lab point the same way
| Metric | What it measures |
|---|---|
| 26% | Share of Anthropic's internal AI R&D work that Claude performed at a "lead" level, as of August 2026 |
| ~30,000 | Number of AI agents simultaneously running research and engineering tasks on Anthropic's main internal platform |
| ~3x | Speedup in a specific dose-response experiment workflow in the MHS proof-of-concept |
| 8 hours | Time it took in one MHS case to build the Claude orchestration layer across multiple lab instruments |
The 26% figure and the 30,000-agent count are internal AI R&D automation metrics; the 3x and 8-hour figures come from a limited lab proof-of-concept. They are not measuring the same population and should not be compared directly.
Even so, the two pieces of news point the same direction. AI first becomes the executor of digital research, then starts taking on part of physical experimentation through robots and instruments. If AI produces better code, better hypotheses, tests those hypotheses faster in the real world, and feeds the results back into itself, the research cycle shortens.
It would be premature to assume the human researcher's role disappears right away. The nearer-term change looks more like an increase in the number and parallelism of experiments one person can supervise. A scientist's value may shift away from time spent moving a pipette by hand and toward deciding what to ask, which results to trust, and which experiments to stop.
This shift is bigger than drug discovery
If AI can reliably operate physical equipment, there is no reason the same structure stays confined to biology. Chemical synthesis, new-materials discovery, battery testing, semiconductor processes, manufacturing quality checks, any industry where hypothesize, execute, measure, revise repeats, has a similar closed loop available to it.
This also changes the frame for sizing AI's economic reach. Until now, the generative AI market has been counted in users, software seats and API tokens. Going forward, equipment, processes and experimental workflows could enter that calculation too. As AI connects to robots, sensors and lab instruments, its total addressable market may widen from automating digital labor to orchestrating physical production and research.
If Tesla is trying to give AI a body to work in the factory through Optimus, Anthropic's direction looks more like attaching hands to AI in the lab. The industries differ, but the common thread is the same. The next round of AI competition will not be decided by better answers alone. It is widening into how safely and repeatably an AI can act in the physical world.
The hardest problems that remain
| Bottleneck | Why it's hard | Signal that would strengthen the case |
|---|---|---|
| Physical throughput | Reagents, incubation time and equipment capacity can't be parallelized infinitely the way software can. | Steady gains in 24-hour autonomous runs, equipment utilization, and throughput per experiment. |
| Reproducibility | A single successful experiment only gains scientific value once other labs and equipment can repeat it. | Improved success rates and reduced variance for the same protocol across multiple institutions. |
| Safety and control | Unlike software bugs, a wrong action in a physical experiment can damage samples, equipment, or raise biosafety issues. | Transparent disclosure of human intervention rates and blocked risky actions, scaling up without incidents. |
| Clinical translation | Good preclinical results can still take years to clear the much higher bar of clinical safety and efficacy. | AI-generated candidates repeatedly advancing to IND filings, clinical entry, and improved real-world success rates. |
Anthropic itself acknowledges this boundary. The company has said it will not run clinical trials directly for now, and is focusing on preclinical work and areas existing pharmaceutical companies underinvest in for economic reasons. As Reuters noted, no drug program's success is guaranteed. Most candidates still fail to clear the much higher wall of clinical safety and efficacy.
Insight Times Editorial Desk
