A board item came back complete. The session log was there, the output looked right. At the bottom of the summary was a handoff note for me: proceed via the enrollment path.
No machine name, no command. Just the phrase, as if I’d know what it meant without any of the context the agent had accumulated.
I did eventually figure it out. I found the right logs and ran the command. But that was all work the agent could have done. The path was in the session. The agent had walked through the setup steps already. That information existed; it just didn’t make it to the place I actually read.
Starting August 23rd, the builder prompt requires a HUMAN-STEP line for any board task that depends on a step I have to run myself. The format is explicit: HUMAN-STEP: <exact commands+paths+machine>. It gets extracted from the full session log and surfaced in a notification when the item needs me. The test is whether I can read the note and run the step without opening anything else.
“Via the enrollment path” is not a handoff. It’s a placeholder that moves work from the agent’s queue into mine, without the context I’d need to close it.
There’s a version of this that looks like good intentions. The agent surfaced the step as human-required and passed it off. That part is right. The failure is in the resolution. At some point I’ll have six items waiting on me across three machines. I’m not going to dig through six session logs to find the commands. The note needs to do that work.
Session notes make sense to the agent that wrote them. In context, “via the enrollment path” carries meaning. But the handoff note isn’t read in that context. It’s read cold, and the person reading it wasn’t there. That’s who the HUMAN-STEP line is for.
What I’d tell myself before shipping that board item: write the HUMAN-STEP line first. Before the summary, before the output. If you can’t write the exact command and machine in full, go back into the session and find what’s missing. The note is the handoff. If the note can’t stand alone, the work isn’t done.