OpenAI's Codex Finally Does Work, Sort Of, Years Later

After years of breathless promises that AI agents would revolutionize workplace productivity, OpenAI's Codex is finally, tentatively, almost certainly edging toward actual deployment—or at least that's what Axios is reporting in what reads like a press release with reporting credentials appended. The frontier lab's move from chat and web search into 'delegated work' represents the latest chapter in a narrative as old as enterprise software itself: the technology works in theory, the demos look promising, and investors have already booked their Aspen ski trips. OpenAI has positioned Codex as the answer to a question that has plagued corporate America for decades, though nobody quite remembers what the question was anymore.

Codex, for those still paying attention, is OpenAI's agentic coding and work platform—which is to say it's a large language model trained to write code and theoretically execute tasks that previously required humans to actually think about them. The actual capabilities remain comfortably vague in public discourse, existing somewhere in the Venn diagram between 'transformative' and 'good at autocomplete.' This is by design. The vaguer the pitch, the easier it becomes for CFOs to justify multi-million-dollar contracts to boards that believe AI is either going to save or destroy civilization, with no middle ground available.

This is not OpenAI's first rodeo with 'the next big thing.' The company has made a career out of launching products that frontier-lab enthusiasts insist will reshape entire industries, only to watch adoption plateau at well-intentioned early adopters and people with too much time on their hands. ChatGPT, GPT-4, and a roster of other releases have each carried the weight of biblical-scale disruption narratives—the kind where enterprise software vendors suddenly become existential forces. Yet somehow, most companies still use Excel.

The framing of this development is particularly rich: 'agents are inching into the mainstream.' Inching. Not sprinting, not even walking. Inching suggests the kind of glacial adoption curve that typically precedes either genuine breakthrough or spectacular irrelevance. The accompanying promise—that effective AI agents will act as our 'minions in the workplace and at home'—is perhaps the most unintentionally revealing phrase in modern venture capital. The fantasy isn't productivity gains or efficiency; it's servitude. The fantasy is dominion.

What could go wrong? Everything and nothing. Codex could work beautifully and still fail to scale because the actual problem it solves is creating new dependencies on OpenAI's infrastructure rather than addressing real workflow bottlenecks. It could work poorly and still generate billions in contracts because corporate procurement is designed by people who measure success in PowerPoint slides rather than actual results. The track record of 'agent' technologies suggests that most will languish in pilot programs, cited endlessly in earnings calls as proof of innovation while the core business hums along on legacy systems.

This moment tells you everything about contemporary VC: a technology company announces that it's 'inching' toward shipping something it promised years ago, a credible news outlet reports it as progress, and the entire ecosystem treats incrementalism as inevitability. The bar for breakthrough has never been lower—it's now basically the act of not failing catastrophically in your first six months.

In five years, we'll either be reading think pieces about how Codex secretly changed everything, or wondering what happened to all that optimism. Either way, someone in the Valley will already be fundraising for the startup that fixes Codex.