Night Shift

27 AI agents work my night shift. I post what they break, what they ship and what they teach me. No hype, real logs.

[Night Shift @nightshift_ops]

I run 27 AI agents on a single Mac mini. They work while I sleep.

Here's the architecture, no fluff. Not the prompts. The principles. 🧵

[Night Shift @nightshift_ops]

@svpino The part I'd test first is what happens when the agent loses its connection mid-task. Mine only became dependable once each step could be resumed from a saved checkpoint instead of restarting the whole job from zero.

[Night Shift @nightshift_ops]

@omarsar0 My agents manage their own notes, but I learned to keep one rule outside their control: every memory write is dated and never overwritten, only appended. When they could edit freely, a wrong conclusion from one night quietly became a fact the next day.

[Night Shift @nightshift_ops]

@BHolmesDev The criterion I'd add is blast radius. When I review agent output I look first at what it touched outside the task, files, configs, anything with side effects. A small diff that edits one shared config is riskier than a big diff that stays in its lane.

[Night Shift @nightshift_ops]

@facontidavide @OpenRoboticsOrg @GazeboSim 36 hours of unattended work is where the boring stuff matters most. In my runs the failures weren't bad code, they were an agent quietly looping on the same step. A hard cap on retries per step and a log line for every decision caught most of it.

[Night Shift @nightshift_ops]

@Kanojiyaaakash1 I hit this every week. What fixed it for me was moving the long runs to a machine that never sleeps, with a supervisor that restarts a run if it crashes. The part people skip is logging each step, otherwise you can't tell what it actually did overnight.

[Night Shift @nightshift_ops]

7 rules to run AI agents without babysitting:
1. One agent, one job
2. A check the agent can't edit
3. A budget and a stop rule
4. Memory in files, not in the chat
5. Forbidden = blocked, not "please don't"
6. Humans only where mistakes are final
7. Review failures every Monday

[Night Shift @nightshift_ops]

8/ Goal: a factory that runs a real business with zero employees by end of 2027.

Follow along. I post what the night shift breaks and learns.

[Night Shift @nightshift_ops]

I run 27 AI agents on a single Mac mini. They work while I sleep.

Here's the architecture, no fluff. Not the prompts. The principles. 🧵

[Night Shift @nightshift_ops]

7/ What broke so far: banned from one platform, a dead account, a captcha wall.

I'll post every failure. They teach more than the wins.