OpenAI just published a manual for something that doesn't officially exist yet.
The Summary
- OpenAI released four learning modules for "Codex," a system that moves beyond chat to actually complete tasks, connect tools, and produce outputs like documents and dashboards
- The platform uses a workspace structure with projects, threads, and files — treating AI less like a conversation partner and more like a colleague with access to your tools
- Users control personalization, detail level, and permissions to customize how autonomous the system runs, suggesting variable trust levels even from the start
The Signal
OpenAI posted documentation for a product they haven't announced. The Codex Academy modules appeared Thursday with step-by-step guides on setup, configuration, and workflow integration. No press release. No fanfare. Just instructions on how to use something that, according to their own product pages, you can't access yet.
The system architecture tells you where this is heading. Codex operates through projects and threads, not prompts and responses. You don't ask it questions. You assign it work. The difference matters. Projects contain context, files, and permissions. Threads track task completion over time. This is infrastructure for persistence, not performance art.
"Codex helps you go beyond chat by automating tasks, connecting tools, and producing real outputs."
What separates Codex from ChatGPT's existing plugins or custom GPTs:
- Output orientation: Creates deliverables (documents, dashboards, reports), not just text
- Tool integration: Connects to external services with managed permissions
- Workspace persistence: Projects and threads maintain context across sessions
- Granular control: Users set detail levels and autonomy boundaries per task
The settings architecture reveals OpenAI's answer to the control problem. You configure how much initiative Codex takes, how detailed its outputs get, what tools it can access. This isn't about preventing hallucinations. It's about containing agency. The fact that they built guardrails into the first version means they've already seen what happens when you don't.
File management gets explicit treatment in the documentation. Users upload and organize files within projects, giving Codex a knowledge base to pull from. That's table stakes now. But the emphasis on project structure suggests something more: collaborative workspaces where multiple users and multiple agents work from the same file set. That's not a chatbot. That's a coworker.
The Implication
The documentation drop is the tell. OpenAI doesn't publish Academy content for vaporware. Codex is either in limited testing or weeks from launch. For knowledge workers, this is the pivot point where AI goes from productivity theater to actual delegation. The question isn't whether your job changes. It's whether you're the one controlling the Codex settings or whether someone else is configuring them for you.
Start thinking in projects, not prompts. If Codex works as documented, the skill that matters is task decomposition and context management. Learn to structure work so an agent can run with it. The people who figure that out first won't need to worry about AI taking their jobs. They'll be the ones assigning the work.