Coscientist: An AI Agent That Translates English into Laboratory Experiments

Gabriel Gomes built an agent that turns plain English into physical experiments, enabling research that humans alone could never sustain. Gabriel Gomes believes the future of chemistry is as much about flasks and fume hoods as it is about code. A chemical engineer at Carnegie Mellon University, Gomes works at the intersection of chemistry and artificial intelligence. His goal is to automate the drudgery of laboratory research, making experiments faster, more accurate and easier to perform. His work has led him to create Coscientist: an intelligent agent that adapts large language models such as GPT-4 for automation and lab infrastructure.

Gomes’s path began in a small town in the Brazilian countryside, where he didn’t own a computer until he was 19 years old. The first in his family to attend university, he found his calling at the Federal University of Rio de Janeiro, when a professor told him, “All of chemistry is inside this Schrödinger equation—all you need to do is solve it!” That idea put him on the path to computational chemistry and, eventually, the White House Office of Science and Technology Policy, where, in 2023–2024, he advised about the risks and rewards of intelligent systems.

Where did the idea for Coscientist come from? It came from a worry. Carnegie Mellon was building a big initiative for an academic cloud lab—$50 million of equipment controlled by a mix of people and robots—and the interface would be via code. I was nervous that my colleagues in chemistry and biology would not use this amazing platform. You basically would have to ask them to think about how they would do experiments in a whole different setting that they are not physically involved in. I like to joke that chemists, organic chemists in particular, have this sense that “my lab is my kingdom, and you shall not trespass.”

My group started in January 2022, and we were trying a few things with large language models and not really succeeding. Then GPT-4 came out on March 14, 2023. I remember one of my students sending screenshots of the white paper on our Slack, and I thought he was pulling my leg—the capabilities were incredible. I remember waking up at 6 A.M. with the realization that “this is how we fix the problem. We can use this to let chemists interface with the cloud lab using natural language.” That’s how Coscientist started.

How has this changed the day-to-day research in your group? There is a before and after for a group like mine. I had a student who joined in 2024, and before joining, he came to me, very nervous, saying, “I’m really interested in all the things the group does, but I don’t have a background in programming.” But because he had access to some of the best state-of-the-art tools and also learned from the language models, he was able to accelerate. In very little time, he learned to do everything. He now does the computational and machine-learning sides of projects that are pushing the boundaries of autonomous scientific research.

Can you describe how Coscientist works? Imagine you want to bake a cake, but you do not know how to measure the ingredients or use the oven. You simply tell Coscientist, “Bake me this chocolate cake.” It figures out the recipe, checks what kind of equipment and ingredients you have to see if it’s possible to bake the cake and gives you the instructions. It can be your guide. You can share photographs and videos and have it troubleshoot the next step.

The very first experiment we did was with our robot, a 96-well plate with food coloring and a target plate. We told the robot, “Draw something cute on the target plate.” Coscientist drew a fish. We don’t know why, but it’s cute. This alone is impressive because ordinarily a human would have to program the robot exactly. From there, we were able to develop one of the largest datasets of experimental chemical reactions with kinetics, which people don’t usually do because of how much work it is. This is the kind of science I believe we’ll be doing going forward: reaching areas we have not touched because of human bias or because the amount of labor was enormous.

Is there anything people should be careful of when conducting research with large language models?