← All clips · Uncategorized

Reins for the agents

PASSING GATE Administrator 12 steps · 37.4s · recorded Aug 30, 11:04 AM · bd2e5e8

Viewing an archived run (20260830T180419Z-bd2e5e8). Back to the current run.

This run has no video.

Speed

Source: e2e/scenarios/test_wf_reins_for_the_agents.py:1 · run 20260830T180419Z-bd2e5e8 · steps carry start_ms/end_ms offsets into the video

Steps

  1. Pause, Restart, Stop and an editor for the prompt each loop runs on.

  2. Shown once; only the hash is stored.

  3. A separate credential for the one-minute cron that is NOT a Claude session — a hung agent polls nothing, so restart cannot be an agent's own verb.

  4. The demo admin password is printed on the site. If the session were the only factor, a stranger could stop the owner's loops — or edit the prompt of an unattended CLI on the owner's machine.

  5. Held in sessionStorage — it dies with the tab. The factor standing between a public admin login and the owner's machine should not outlive the window it was typed into.

  6. Reported by the loop on its heartbeat — NOT read off this server's disk. Plan pushes are [skip ci], so the deployed checkout can be days behind the file the loop actually read.

  7. Confirmed before it is sent: this is code being handed to an unattended loop on the owner's machine.

  8. update_prompt · claimed by agent · done. A LIVE loop applies its own prompt edit, so the running context and the file never diverge.

  9. The platform clock moved to 2026-08-30T20:04:21Z and the scheduler ticked.

  10. One card is what the agent last said about itself. The other is what something outside it can see. The disagreement is the diagnosis.

  11. Confirmed first: whatever the loop is part-way through is abandoned.

  12. What took an SSH session and a morning of confusion on 2026-08-23 is two taps and one cron tick.

Notes

Sign in to leave a timestamped review note. Notes land in the same review backlog as in-page annotations.

  1. No notes on this clip yet.
Run history (5)