Night Shift Agentic Workflow

Takeaways

• Day shift / night shift split: Human time is for specs, architecture, and thinking; agents run autonomously overnight on completed specs while you sleep.
• Specs are for you, not the agent: Detailed ./Specs docs organize your thinking (edge cases, design); draft-* specs are ignored. Small @AGENTS.md router points agents to workflow/docs/skills.
• Night loop is heavily automated: Prep (clean tree, fix tests) → pick bugs/features → test-first → multi-persona review agents (REVIEW_PERSONAS.md) → implement with strict lint/types → full suite → commit with changelog → loop until done.
• Morning review, not babysitting: Review stacked commits, fix docs/workflow when agents err (postmortem the process, not just the code), manual-test for gaps in understanding.
• Feedback loop is the product: Burn tokens on validations and review steps so humans never catch obvious issues; constant doc/workflow improvement amortizes over the project.
• Results: ~5× faster, better quality, less context-switching, more fun — agents idle during the day, human idle at night.

Notes

Jamon Holmgren’s answer to agentic burnout: treat human attention as scarce and agent tokens as abundant. Key anti-patterns avoided: reading agent plans, prompt-and-reprompt loops, babysitting idle agents. Uses @AGENT_LOOP.md to kick off overnight work with Claude Code, Cursor, or Codex. Companion piece: The eight best ways I’ve improved my AI agent’s code.

Open questions

• How portable is the multi-persona review loop (REVIEW_PERSONAS.md) to smaller codebases or solo projects?
• Stacked commits + stacked PRs: best practice for teams vs. solo?