AutoProber made HN ask what happens when an AI agent moves real hardware

AutoProber got HN’s attention because it looks like the kind of hardware project that could become a résumé in one repo: an AI-driven flying probe stack built around a CNC machine, microscope, oscilloscope, dashboard, and a lot of hacker pragmatism. The post reached more than 220 score and 40 comments, but the community’s reaction was not only admiration. It was also a demand for proof.

The GitHub README describes AutoProber as a hardware hacker’s flying probe automation stack. The intended workflow is that an agent ingests the project, verifies hardware, runs homing and calibration, identifies a new target on the plate, captures microscope frames with XYZ positions, labels pads, pins, chips, and other features, then adds probe targets to a web dashboard for human approval. Only approved targets are probed.

The prototype stack uses a GRBL-compatible 3018-style CNC controller, a USB microscope, a Siglent oscilloscope over LAN/SCPI, an optical endstop, and an optional network-controlled outlet. The repo includes Python control code, a Flask dashboard, CAD files, and operating documentation. The most important part may be the safety section: the project explicitly treats itself as a machine-control system, not a normal web app.

That safety model is specific. The CNC probe pin is not treated as a trusted endstop. An independent safety endstop is read from oscilloscope Channel 4, which must be monitored during motion. Channel 4 triggers, ambiguous voltage, CNC alarms, and real X/Y/Z limit pins are stop conditions. Recovery motion is not automatic.

HN’s skepticism focused on the gap between a cool demo and a reliable instrument. Commenters asked what the AI actually does: finding pins, comparing against a SPICE model, reverse-engineering circuits, or just annotating microscope images. Others worried that real PCB imaging, fiducial math, and sub-millimeter accuracy are hard enough before an agent is allowed to move a physical probe. A small positioning error can damage a board.

That is why this thread matters. AutoProber is interesting not because it magically solves hardware hacking, but because it exposes the design problem every physical AI agent will face. Once software can move metal, trust shifts from prompt quality to interlocks, calibration, bounded motion, operator review, and evidence that the machine knows where it is.