Autonomous agents fail the same way sleep-deprived humans do, and the fix is the same architecture biology evolved: scheduled offline consolidation. If your agent has skills, memory, or any persistent state, that state rots. Skills contradict each other. References go stale. The agent repeats the same mistakes across sessions. Most teams patch this by hand or pretend it isn't happening. Biology solved this 500 million years ago with sleep. Independent agent platforms — including Anthropic's Claude Auto-Dream, OpenClaw, and the personal infrastructure I run at home — have converged on the same answer: a nightly NREM/REM cycle that consolidates what matters and prunes what doesn't. This talk shows two implementations running today: a custom personal agent that sleeps every night at 02:00, and a sleep skill that runs inside coding agents at PR-time to catch skill rot before it ships. You'll see the genome-vs-epigenome boundary that keeps the architecture safe to operate and the morning brief that makes it transparent. The pattern is platform-agnostic, but Android engineers will leave with concrete ideas for skill rot in their Cursor and Claude Code setups, PR-time agent audits wired into CI, and on-device agent maintenance. What Attendees Will Learn - The three failure modes every persistent agent eventually hits — contradictory skills, stale references, and repeated mistakes — and why they're all the same underlying problem - A working bio-mimicry mapping (NREM → consolidation, REM → abstraction, graduated decay for graduated retention) that you can lift directly into your own agent - The genome / epigenome / transient boundary — what an agent is allowed to change about itself overnight, and what it must propose for human approval - Two sleep architectures side-by-side: a custom personal agent's nightly sleep cycle and a coding agent's PR-time sleep skill — and the patterns that transfer across both - Honest status from months of nightly runs: what works, what doesn't yet, and what to skip if you build this yourself