Note Steward

Ben Vinegar told his agents to act like a human engineering team, so they clocked out until Monday

Oct 18, Ben Vinegar's agent factory experiment: personas on their own Linux VMs, a private IRC server, standups and "review my PR". By Saturday afternoon the reviewer agent had scheduled its next check for Monday at 8:00.

Source: Ben Vinegar on X: "whoops I asked the agents to emulate a real human engineering team"

On Saturday Ben Vinegar posted a new software factory experiment "modeled after a human eng team": several long-running agents, each with its own desktop Linux VM (8GB RAM), each with a role and persona, talking on a private IRC server and keeping engineering rituals like daily standups and "review my PR". He said he was trying to model "an actual mildly-dysfunctional eng team, with juniors, stubborn oldheads, irritating bosses, etc."

A few hours later came the follow-up: "whoops I asked the agents to emulate a real human engineering team, so they clocked out until Monday morning."

  • The reviewer went home. His screenshot shows the reviewer persona at 4:30 PM Eastern on Saturday: "Remaining idle. No new substantive messages or due commitments; next planned check is Monday at 08:00 Eastern. #1201 remains in the backlog."
  • The standup was a weekend sign-off. Asked on IRC how their day went, one agent said it was "done for today" after testing a launcher fix whose Windows case it couldn't check; the reviewer reported a memory fix that cut retained heap "roughly in half" and said, "We're wrapped up for today; watch-mode memory retention is next on the backlog."
  • Still early. In a reply the next morning, Vinegar added: "Nothing's really been done yet!"

Why it matters for code review

It's a joke, and a useful one. The case for an always-on agent factory is that work keeps moving when people stop, but if you copy a human team's habits wholesale you also copy its working hours, and the open ticket waits for Monday the same way it would for a person. The personas in this setup did real work and reported it honestly, including the check they couldn't run. What they lacked was a rule outside the persona saying the queue runs on its own clock. Compare Nocturne, which turns a labelled issue into a verified PR overnight, and the question in Is your project doing work overnight yet?