OpenAI Codex CLI powers 2.5-hour feature build

The real story isn't that Codex can write code; it's that the harness around it can keep an agent productive for hours without a babysitter. Worktree isolation and tight feedback loops turn the demo into a believable dev workflow.

• –Worktrees are the safety valve: the agent can branch, fail, and retry without trampling the main repo.
• –Long autonomous runs matter more than one-shot code generation because real feature work is mostly test, debug, and revise.
• –This shifts human effort up the stack toward specs, guardrails, and review instead of line-by-line implementation.
• –OpenAI is clearly betting on agent-first development, where the repo, docs, and tests do the heavy lifting.
• –For teams comparing Codex with Claude Code or Copilot CLI, durability and orchestration are the interesting battleground now.