This Week's Term: Self-Driving Lab. A research operating model in which AI proposes the next experiment, robotic instruments run it, ML models read the result, and the loop runs again, with light human supervision rather than a human at every step.
The term has been in scientific use since at least 2018, championed publicly by Alán Aspuru-Guzik at the University of Toronto and through the Acceleration Consortium. What has changed by 2026 is that the lab-in-a-box version has moved from research-paper claim to routine operating mode in a handful of places, and that it now interlocks with multi-agent systems in a way that was not possible five years ago.
Where it actually runs
A few worked examples, from 2024 to 2026:
- The Acceleration Consortium now coordinates roughly fifty autonomous robots across member institutions.
- Northwestern, February 2026. Ewen Callaway covered Meagan Olsen in Nature news: a GPT-5-driven autonomous biology setup that beat the human state-of-the-art on cost-reducing reagent swaps for cell-free protein synthesis.
- Argonne's Polybot runs autonomous polymer-formulation experiments at high throughput.
- Commercial side, Q1 2026. Automata closed a $45M Series C. Chemspeed/SciY launched a vendor-agnostic SDL platform at SLAS2026. Ginkgo Bioworks opened a web interface to its Cloud Lab.
The pattern across all of these is the same. AI proposes. Robots execute. ML interprets. Humans set the question and intervene where their judgement is irreducible. That last clause is exactly the division of labour I unpack in this issue's framework section on the Stanford Virtual Lab. The Virtual Lab automates the meetings. The self-driving lab automates the wet bench. Wire them together and you have something closer to an end-to-end agentic research stack.
Why it matters in 2026
Jason Kelly is the co-founder and CEO of Ginkgo Bioworks and the former chair of the US National Security Commission on Emerging Biotechnology. He has the clearest framing I have heard this year for why self-driving labs are not just a niche scientific story.
Computers, he notes, transformed industries made of information — media, telecom, finance. They did not really transform industries made of atoms — building materials, food, fertilizer. The interesting thing about a cell, he argues, is that it is also programmable, and what it programs is matter. "Cells are the only thing we have that are programmable in the physical world." Self-driving labs become the operating model for programming the physical world at scale.
Kelly's picture of what becomes possible is a single scientist directing thousands of agentic scientists, each pushing down an experimental line through an automated lab and reporting back. The reason he thinks this might work: most discovery is not the result of unusually brilliant thinking; it is the result of trying things no one has yet tried. "If the model can try things with an automated lab, why can't it discover something new? Most of discovery is trying new things."
I do not think the timeline is as short as Kelly implies. Yet, self-driving labs are not only a research curiosity. They are the operating model to bet on when you decide programming atoms is more important than programming bits.
This is also where the Project Hail Mary metaphor from this issue's leadership section gets uncomfortably literal. Rocky has many hands. The self-driving lab is many hands. The question of what work we put them to is no longer hypothetical.
Real-world implication
For business leaders outside biotech, three questions are worth running this quarter.
- Is there a part of your operation that consists of repeatedly proposing, running, and reading the results of a structured experiment? Marketing creative testing, supply chain optimisation, retail pricing, drug formulation, materials science, even financial backtesting all qualify.
- Could the proposing step be done by an agent, the running step by some form of automated execution, and the reading step by an ML model? If yes, you have the rough shape of a self-driving system.
- Where is human judgement irreducible in that loop, and is your organisation currently spending the right amount of senior time on that part?
Most leaders find at least one candidate process when they run those three questions seriously.
Your action step
Sit with one cross-functional team this week and map a single experimental loop in your organisation against the propose / run / read / iterate structure. Do not try to automate it on the spot. Just label which steps a current AI tool could do well, which steps require a human today, and which steps are limited by physical execution. The map is the deliverable. The conversation that follows the map is where the actual strategy lives.
If you want a faster read on which loops in your business look most like a self-driving lab in waiting, that is the kind of mapping I do in AI strategy advisory sessions and working backwards workshops with leadership teams.
Sources
- Alán Aspuru-Guzik, University of Toronto, and the Acceleration Consortium, originating use of the term since 2018.
- Ewen Callaway, Nature news, February 2026, on Meagan Olsen's GPT-5-driven autonomous biology setup at Northwestern.
- Argonne National Laboratory, Polybot autonomous polymer experiments.
- Jason Kelly, Ginkgo Bioworks, public talks on programming the physical world.
- SLAS2026, Chemspeed/SciY vendor-agnostic SDL platform launch.