My AI Adoption Journey

· ai-agents coding · Source ↗

TLDR

  • Mitchell Hashimoto documents a 6-step progression from chatbot skeptic to running background agents daily for real coding work.

Key Takeaways

  • Chatbots fail for brownfield coding; value only emerges when using agents with file read, program execution, and HTTP access.
  • Forcing himself to reproduce manual commits with Claude Code built genuine expertise in what agents handle well vs. poorly.
  • End-of-day agents doing deep research, parallel prototyping, and GitHub issue triage (read-only, no auto-responses) provided a productive “warm start” each morning.
  • “Harness engineering” means updating AGENTS.md and building programmatic tools each time an agent makes a mistake, preventing recurrence.
  • Goal is an agent running at all times; current reality is 10-20% of a working day with a single agent, not parallel fleets.

Why It Matters

  • The step-by-step framework is derived from first principles, not theory: each phase was validated by doing work twice and measuring friction directly.
  • Turning off agent desktop notifications and controlling interruptions manually is identified as a concrete prerequisite for staying in deep work.
  • The Ghostty macOS command palette, shipped to users, originated from a single Gemini screenshot prompt, grounding the abstract journey in a real shipped artifact.

Mitchell Hashimoto · 2026-02-05 · Read the original

Related coverage