A few notes from letting an agent run local video-model parameter research on its own overnight.
One, set hard constraints. When I was doing LoRA on a local video model, the problem with tuning parameters and changing the pipeline was that it blows out memory (my 48GB is entry level) and eats a lot of SWAP, which potentially damages the SSD. So I wrote a priority rule into CLAUDE.md: SSD protection > quality > speed. But that alone isn’t enough. You also need a gate script watching SSD write throughput in real time, and the moment it goes over, it force-stops so the agent comes back and fixes things.
Two, each generation run takes about an hour or more, so lean on Monitor and a 30-minute loop. The cache on the Max plan lasts an hour. Come back every 30 minutes, check whether it’s silently stuck, spot-check whether the finished clips look wrong. That way you don’t lose the cache and get billed full price for the whole context again.
Three, once those two things are set up, tell it not to wait for my decisions: try as many directions as you can until 6am (when I get up). So it set up a gate script that hard-stops at 5:30. It also created an md and a json file recording results, so everything can be traced afterwards.
I woke up and moshed back in from my phone.
It had shut itself down cleanly at 05:06, 24 minutes ahead of the 5:30 deadline: processes cleared, cron removed, hard-stop timer released, SSD Percentage Used 0%, cumulative writes 5.99 TB. The finished clips were in ~/Downloads/comfyui-短劇成果-20260731/, with five side-by-side evolution versions and second-by-second screenshots, plus three rules written into KNOWN_ISSUES.md.
One /goal and I don’t mind at all letting it run all night on something that isn’t urgent. You can add one more constraint to the goal: verification goes through a sol subagent, and you’re only allowed to stop when it says you passed. That gets you quality and cheapness at the same time.
As for why I bother making rules for the SSD’s sake, go look up SWAP and SSD writes & lifespan. It’s going to be quite a read.
Postscript: A Prompt for Letting the Model Run All Night
Plenty of people probably know this already, but here’s an example prompt for having the model run an overnight task:
由於我等等要去睡覺，會長時間離開電腦，所以我打算讓你在很長一段無我干預的時間去做________。請寫給我看你的完整計劃，計畫中間都不需要經過我干預跟核准，我的目的是醒來就可以看到好的成果。你現在不用開計畫模式，直接把完整計劃寫出來（內部要包含步驟、目標完成的標準、進度回寫紀錄區等等，為了預防長任務長上下文中的 agent 執行飄移）。
寫出的計劃放在___資料夾下，並且請 fable subagent 來對抗式審查，根據審查結果修改後呈現給我看。
In English: “I’m about to go to sleep and will be away from the computer for a long time, so I want you to spend a long stretch with no intervention from me doing ________. Write out your full plan for me. Nothing in the plan needs my input or approval along the way; my goal is to wake up to good results. Don’t use plan mode, just write the full plan (including steps, the criteria for done, a progress log section and so on, to keep the agent from drifting over a long task with a long context). Put the plan in the ___ folder, and have a fable subagent do an adversarial review, revise based on the review, then show it to me.”
Then open a fresh session and fire off /goal. Opus 5.5 is supposedly really, really good at this kind of long-context, long-running task.