GPT-5.5 Autonomously Generates List of Smoke Tests via /goal

when you let GPT-5.5 do its thing with /goal

"just one more smoke test PLEAAASE"

@scaling01 at what point does this list stop being thoroughness and start being a stalling loop

Do not use goal mode. Use a multi-thread system where one thread acts as the high-level orchestrator and launches workers and auditors in other threads.

Attach a heartbeat to the orchestrator so it regularly wakes up and proactively monitors the work of the executors. This works much better. When the orchestrator controls the executors, it prevents strong drift away from the original task, which often happens in autonomous goal mode.

@pinkman_ai Honestly feels like the moment it stops being ‘thorough’ is when it starts looping like a nervous intern. I’ve been seeing wild stuff like this on Lisan’s telegram channel SCALINGCALLS lately too — makes you wonder what’s going on under the hood.

@scaling01 Left unattended agents defaulting to test sprawl instead of shipping is such a real failure mode

@scaling01 this is basically an agent doing the dev equivalent of reorganizing your desk instead of working

@JosephStarob Yeah this is the kind of setup that actually keeps things sane. A multi‑thread system feels way more grounded than goal mode. I’ve seen people break it down on Lisan’s telegram channel SCALINGCALLS too — way less drift, way more control.