Sleepwalker - Overnight Agent Fleet

A fleet of AI coding agents that work overnight across local and cloud runtimes, then wait in one Morning Queue for a single swipe-through review.

Role: Designer and engineer. Team: Solo. Status: Complete. Year: 2026.

Overview

Sleepwalker runs a curated fleet of AI coding agents while I sleep and hands me one place to review what they did when I wake up.

Overnight agents have three recurring failure modes: surprise bills from runaway loops, false "done" claims, and no clear answer to "what did it actually do?". Sleepwalker treats all three as design constraints. Every irreversible action is parked for approval, every run has a budget cap, and everything lands in a single Morning Queue with a timestamped audit log.

How it works:

  • Local fleet members run as scheduled Claude Code tasks on the Mac, each with its own permission scope.
  • Cloud fleet members run as Claude Code Routines and only ever open pull requests on dedicated branches.
  • Codex and Gemini CLI agents can be scheduled from the same dashboard.
  • Safety hooks cap tokens per agent, defer irreversible actions, and write an audit trail.
  • The Morning Queue merges local actions and cloud pull requests into one swipe-through review, color coded by how reversible each action is.

I built it solo over about 230 commits.

Key features

  • One dashboard to author, schedule, and deploy agents across four runtimes
  • Morning Queue: a single swipe-through review of everything that ran overnight
  • Irreversible actions deferred until a human approves them
  • Per-agent token budgets enforced by hooks
  • Cloud agents work only through pull requests on dedicated branches
  • Timestamped audit log with reversibility color coding

Technical details

  • Dashboard: Next.js 15, React 19, TypeScript, Framer Motion
  • Agents: Claude Code scheduled tasks and Routines, Codex, Gemini CLI
  • Safety: Claude Code hooks for budgets, deferral, and auditing
  • Testing: Vitest

Challenges

  • Making autonomous agents safe to leave running without supervision
  • Unifying local tasks and cloud pull requests into one review surface
  • Catching false success claims before they reach the morning review

Tags: AI Agents, Claude Code, Next.js, TypeScript, React