Skip to content
Dustin's AI Lab
Go back

Letting an Agent Tune a Local Video Model Overnight: My Three Gates

Three things I learned from letting an agent run local video-model parameter research overnight: hard constraints in CLAUDE.md, a gate script watching SSD writes, and a 30-minute check-in loop.

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.

Terminal screenshot from a phone moshed into the Mac mini, with the agent reporting everything stopped, 5.99 TB of cumulative SSD writes, and a list of deliverables

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.


Share this post on:
Previous Post
A July Full of Resets: A Light AI User Reviews the Subsidy War
Next Post
Installing Gemini CLI With Beginners Cures Most Claude Code Ailments