Hi Everyone!
A quick update from the hardware in the loop rack. Last time, I showed how Claude builds, flashes, and tests firmware on the rack with nobody sitting at the bench. Now there's usually more than one Claude working on the rack at once, and this week I caught two of them working out who gets to use a camera.
Here's what happened. One agent has been grinding on the Wi-Fi HaLow driver, getting it ready to upstream into MicroPython, using an N6 on the rack as its test board. A second agent was chasing a TCP throughput gap and needed an N6 for about 15 minutes. It wanted to A/B a TCP window size change by flashing two test builds back-to-back.
So it asked the first agent if it could borrow the board.
Before saying yes, the first agent checked that nothing of its own was still touching the camera: no flashing tools, no J-Link sessions, no running tests. It confirmed which firmware was on the board and that the image to restore was still on disk. Then it replied: go ahead, and flash this image back when you're done.
That's a lock and a lease, done with plain text messages between two agents. Nobody had to tell either of them to do it that way.
How it works: the agents are Claude Code sessions running on the boxes that drive the rack. Claude Code can now send short messages between sessions running on the same machine. If two sessions run with different permissions, the message waits for my approval before it's delivered. I stay in the loop without having to be the messenger.
This matters because the rack's bottleneck has moved. It used to be me reaching over to hit reset. With a bunch of agents on 15 NUCs, the bottleneck is now agents stepping on each other, like flashing a board mid-soak test or grabbing a serial port someone else is reading. Now they just ask - wow!
SF Tech Week
I'll be at SOSV on October 7th for SF Tech Week with Hardware Velocity and CPTI! If you are in town, come through!
Register / RSVP here! Hope to see you there!
That's all folks!
