Promoting AI agents

TLDR

• DHH argues coding agents have crossed into supervised production use in 2025, with models like Claude Opus 4.5 and Codex 5 delivering real results.

Key Takeaways

• Agents gained practical power not from smarter models alone but from tool access: terminal control, test running, web search, and service calls.
• DHH uses OpenCode, a terminal harness that lets you switch between models, capture sessions, and theme-match with Omarchy.
• He rejects the inline autocomplete model (Copilot/Cursor style) in favor of autonomous agents he reviews after the fact, more like managing a team.
• Claims of 90%+ AI-written code are not his experience; supervised collaboration on bugs, features, and drafts is what he can actually vouch for.
• He names Claude Opus 4.5, Codex 5, Gemini 3, MiniMax M2.1, and GLM-4.7 as the current capable generation worth trying.

Why It Matters

• The shift from autocomplete to autonomous agents changes the human role from co-typist to reviewer, which is a different workflow and skill set.
• OpenCode becoming a default in the next Omarchy version is a concrete distribution signal for terminal-based agent tooling.
• The honest ceiling DHH describes – supervised collaboration, not vibe coding – sets a more useful baseline than the 90% claims circulating online.

DHH · 2026-01-07 · Read the original