How Codex Happened: inside OpenAI's coding-agent comeback
At a glance
Oct 15 Every deep dive (Shipper + six OpenAI interviews): Cloud flop → Rust local agent → no-editor desktop app → ChatGPT merge; learnable arc on form-factor bets, separate model lines, and agent-native UX. This workflow is listed under product, coding-agents, learnable-paradigm, openai, case-study.
Agents / tools
Codex, ChatGPT desktop, GPT-5.x Codex, Claude Code
SuperCompress v2 compresses context against the query (keeps answer-critical lines verbatim): ~64% mean cut on their B5 coding-agent suite with 24/24 evidence-retention passes; MCP/plugin for Cursor/Claude Code/Codex plus hosted API and open weights.
Empirical study: 34% of popular active repos ship agent instruction files (AGENTS.md dominant); half exceed ~100 lines; ~3/5 rules could be script-checked but <1/20 name the check—actionable lint: short map, name enforcers, one rulebook.
Magic Patterns’ MagicOS is an infinite canvas for PRDs, GitHub context, screenshots, and side-by-side UI variants; pick a winner and hand it to Claude Code, Cursor, Codex, or any MCP-compatible coding agent instead of losing the design journey in a linear chat log.
junegunn (fzf author) documents how shell CTRL-T bindings die once you live in an agent prompt—and how tmux PREFIX bindings restore fzf: tmux-fzf-complete (files/dirs/URLs/on-screen paths/sentences) and fzf-git.sh (commits, branches, PRs) inject selections as keystrokes into Claude Code or Codex. Requires tmux 3.7+ floating panes and agent settings that keep scrollback (/tui default for Claude Code; Codex alternate_screen=never). Pattern: make agent prompts as completable as the shell.