A senior engineer shares their evolving AI-assisted development workflow, structured around three shifts: a Day Shift for spec creation with Claude/Opus, a Night Shift for autonomous multi-agent implementation using OpenAI Codex, and a Morning Shift for human review. The Night Shift orchestrator breaks specs into PRDs, spawns drafting, reviewing, and implementing sub-agents in a quality-gated loop, committing only after tests and code reviews pass. The author emphasizes that quality remains non-negotiable and outlines how known LLM weaknesses (lack of self-reflection, training bias, sycophancy) are mitigated through externalized context, fresh review sessions, and serialized writes.
Table of contents
How I Got HereWhat Agents Are, and Aren’t, Good AtThree ShiftsThe Day ShiftGet Max Kuklin’s stories in your inboxThe Night ShiftThe Morning ShiftInside the Night Shift OrchestratorConclusion40.3K Impressions