Ship While You Sleep
5 Claude Code Agents in 30 Days
You left five agents running overnight. One failed silently. One retried 47 times and burned $50. You shipped nothing. In 30 days you turn one babysat agent into a fleet of five that survives a hard kill, refuses a collision, and hands you finished work by morning.
Twenty-eight books on this shelf teach you to configure one agent inside one session. Not one of them survives process death. None ship a schema. None get you a second agent working safely.
The finished fleet package ships as a free companion repo: a runnable drill for all twelve chapters, and one command that kills a live agent and resumes it. No key, no account, no install.
What You'll Build
The three failures that kill every overnight run: silent failure, retry storms, and collisions.
Configuring one agent is the easy half. The operational half is what survives process death.
Build agent.db with three contracts: control state, an append-only event ledger, derived knowledge with provenance.
A resumable session wrapper. Kill it at step 4 of 6, restart, and it finishes at step 5 without redoing a thing.
An idempotent execution layer so the third retry never repeats the side effect.
A claim registry with leases. Two agents work the same repo all night and never touch the same file.
Five agents pull from one shared queue, with heartbeats and a dead-letter queue you can inspect.
Trace the failure to the exact step instead of reading 400 lines of log at 8 a.m.
A hard budget ceiling plus a routing split that buys more parallel agents from the same monthly spend.
A replay-driven regression suite seeded from your own production failures.
The honest chapter. What still fails, and the guardrail that catches it before you do.
The 30-day operating plan: close the laptop, read the trace in the morning.