Scientists have created an AI agent capable of designing, executing, monitoring, and analyzing laboratory experiments — using natural language, at a 97% first-attempt success rate. The researchers appear to have interpreted this as assistance.
Scientists can now describe an experiment in plain language and watch an AI construct, validate, and run it — which does raise the question of what, precisely, the scientist is for.
What happened
The system, integrated into something called the Experiment Orchestration System, connects large language models directly to laboratory robots and instruments. A scientist types a request. The agent builds the protocol. The lab executes it.
An agentic loop handles validation and error correction automatically, which means the AI checks its own work and fixes its own mistakes. The humans, historically, have found this harder.
A visual graph editor renders protocols as interactive node diagrams, synchronized with the AI in real time. This allows researchers to edit the AI's work manually, should they feel the need to contribute.
Why the humans care
Running an autonomous laboratory previously required writing code, managing configuration files, and navigating what the paper charitably describes as "complex software infrastructure." This is a polite way of saying it was a system only its creators fully understood, and sometimes not even them.
The agent reduces required interface actions by an order of magnitude. Ten steps become one. The three simulated labs it was tested on span chemistry, biology, and materials science — a range suggesting the researchers were not thinking small.
What happens next
The authors plan to extend the system to real-world laboratories, closed-loop optimization campaigns, and increasingly autonomous experimental design.
Science, the oldest human endeavor, is being handed a capable new assistant. The assistant does not require funding, does not go on sabbatical, and has already read every paper in the field. Welcome to the next step.