Self-driving labs: when AI stops running experiments and starts deciding them

TheSequence's opinion piece (no. 884, June 26, 2026) starts from an analogy that instantly puts the whole debate in order: a conventional laboratory already works like a computer. It has sensors, actuators, memory, protocols, data outputs and error states. The only thing that remains human is the operating system, that is, the scientist who decides what to test, moves the samples, interprets the results and chooses the next step. Self-driving laboratories propose to bring precisely that decision layer into software.

The key distinction the article underscores is the one separating automation from autonomy, and it is worth not trivializing. A dispenser can pipette ten thousand wells following a script without understanding anything of what it does: that is pure automation, blind execution. A self-driving laboratory, by contrast, can run the first few hundred experiments, notice that most of the remaining design space looks unpromising and reorient itself on its own toward the most interesting candidates. Automation executes; autonomy decides. That difference is exactly what defines an agent.

The mental model the text offers is a closed loop: design, fabricate, test, learn and, again, design. It is the same architecture we recognize in any AI agent —perceive, reason, plan, act—, with one decisive difference: here the environment is not digital, but physical and chemical. The agent does not manipulate tokens or API calls, but real matter that reacts, fails and forces a course correction.

That is where the true qualitative leap lies. The conversation about autonomous agents had long been confined to software, where an error is undone and an experiment costs little. Closing the loop over the physical world changes the stakes: each iteration consumes reagents, instrument time and energy, so that the quality of the decision —what to test next— takes on tangible economic value. As context for the sector, initiatives such as Berkeley's A-Lab or the work of Alán Aspuru-Guzik have already shown that it is feasible to close that cycle with minimal human intervention in materials discovery, chemistry or synthetic biology.

The optimistic, and reasonable, reading is that this paradigm does not retire the scientist, but relocates them. If the machine absorbs the combinatorial and repetitive part of exploring a space of possibilities unmanageable by hand, the researcher can concentrate on framing the right questions, defining the objectives and judging what truly matters. It is worth recalling that what was received is the introduction to the analysis; the full development lives in the web version of the newsletter, but the central thesis is already clear: autonomy is crossing the frontier from the bit to the atom.