Paper
Unmodeled states and uncertain action outcomes in agentic scanning tunneling microscopy
Summary
Conditioning an STM tip has traditionally been hands-on expert work. quailbot puts a frontier LLM agent inside the microscope's feedback loop, so the agent sees the instrument's response to every change it makes. After a short apprenticeship, the agent conditioned tips end to end on a real STM and passed independent verification.
Media
Paper
Approach
quailbot connects the agent directly to the microscope. The agent reads and sets instrument parameters, runs conditioning actions, and gets the instrument's readback after every change. State-changing commands stay locked until the operator enables them, and every call goes into an evidence log.
The agent learned the job the way a new lab member does. An expert showed it tip shaking and bias pulsing, and taught it to tell a dirty surface from a bad tip. Then the expert let it work alone.
Outcome
Two frontier models delivered measurement-ready tips from the launch prompt alone, with no further human help. The first of these runs finished in 70.8 minutes. One delivery came on a microscope whose coarse walker was jammed. Harder cases show where autonomy stops today. When the operator secretly loaded a bent tip, the agent noticed that something was wrong but could not find the cause. The next step is agents that infer hidden and unmodeled experimental states and track what each action actually changed.