Should a lab aim to be lights-out?
Usually not. A fully unattended lab makes sense for a narrow, stable, high-volume process that rarely changes. Most research is the opposite: protocols evolve weekly, samples surprise you, and the important decisions need a person who understands the science. The better goal is automation that runs the repetitive steps reliably and hands decisions back to people at the right moments.
Seven principles
People keep the judgment
Machines execute and agents suggest. Decisions about what to run, what to trust and what to try next stay with the scientist.
Automate the tedious, not the interesting
Plate moves, transfers, data entry and reformatting belong to machines. Experimental design and interpretation belong to people, with good tools.
Design every handoff
Between a person, a robot and an agent, each handoff should be explicit: who acts next, with what information, and what happens if they do not.
Make the system legible
Anyone at the bench should be able to see what is running, why, and what needs their attention, without reading logs.
Agents propose, people approve
AI agents are useful for planning experiments, writing transfer files and summarizing results. Their work is shown for review before it touches a sample.
Build on what the lab already has
Connect existing instruments, Benchling, schedulers and LIMS rather than replacing them. The best integration is one nobody has to think about.
Built to last
Coast redwoods live for centuries. Lab software should be understandable, maintainable, and handed over so the team can own it.