Notes from running a production multi-agent AI OS , agents, local LLMs on GB10, observability, and what actually holds up.
A production Hermes self-evolution loop logged 2,450 identical failures over 49 days because a monitor process held stale code in memory. A twice-daily audit agent caught the pattern and forced a restart.
What actually breaks when a multi-agent stack runs for weeks with no one watching, and how a build log plus dashboard keeps it fixable.