GoodBarber CTO reveals three checks for multi-agent loops after 964 hours of AI work

GoodBarber CTO Dominique Siacci published a post-mortem of a six-day autonomous refactor: 28 agents, 964 hours of session time, over 6,000 commits and 717 deployments. Five failures never turned the build red, so he proposed three checks: session heartbeats, machine-checkable rules and scope audits.

• 28 agents, up to 20 at once, 964 hours of sessions and 6,000+ commits in six days
• Median integration wait grew from under a minute to 17.5 minutes as workers rose from 4 to 20
• One worker went silent for nearly three days: heartbeats tracked units, not sessions
• An audit found 516 controls left on the old path with no recorded decision