“I'd been finding lately (particularly with Opus 5.5) that there was a ton of drift and instances of subagents not doing anything and it was evading the attention of the TL despite the instructions I gave it to have them report in at a regular cadence.”

Tracon tails transcripts in ~/.claude/projects and checks each worktree with ps, lsof and git, then flags: blocked on a prompt, idle, looping on a failing command, simultaneous edits to files other subagents are editing, and drifting off the brief. A Haiku judge classifies; pressing a key sends the lead a note to decide. The finding is that the orchestrating model's own check-ins ('report at a regular cadence') are not a reliable supervision channel, so supervision moves outside the agent loop. Same shape as the machine-level guards in the Code by Ramiz piece. 1 point on HN, so weak social proof; the claim is first-hand.