Promoting AI agents

Takeaways

• DHH’s stance shift: agents “came alive” when given tools beyond reasoning — terminal, tests, web search, Agent Skills in plain English.
• Terminal harness > in-editor autocomplete: autonomous agents feel like a team you review; Copilot/Cursor-style completion “steals the keyboard” mid-thought.
• Promotion granted: current models (Opus 4.5, Codex 5, Gemini 3, MiniMax M2.1, GLM-4.7) can produce production-grade contributions to real codebases.
• Not pure vibe coding: supervised collaboration is here today; 90%+ agent-written code claims don’t match DHH’s quality bar.
• Tooling: uses OpenCode (multi-model terminal harness); becoming default in Omarchy next version.
• Call to action: try real work on Opus et al. via OpenCode — “most exciting thing since the internet.”

Notes

HEY World post (Feb 2026). Bridges Lex Fridman podcast sentiment (anti-autocomplete) with new agent paradigm. Acknowledges hype vs. profession uncertainty but argues engineers shouldn’t sit this out.

Open questions

• What % of Basecamp/HEY code is agent-written under DHH’s supervised model?
• OpenCode vs. Claude Code vs. Codex for Rails/monolith workflows?