OpenAI's planned Ona acquisition gives Codex a room to work in. OpenAI says more than 5 million people use Codex each week, up 400% from earlier this year, and that Ona has helped 2 million developers work in secure, reproducible cloud environments. [1]

The paper's June 14 feature said OpenAI's buying rails were public before its S-1. Ona narrows that argument from the IPO frame to the operating frame. A long-running coding agent is not only a model. It is a place, a permission set, a log, and a reviewer.

OpenAI's announcement says Codex work is increasingly unfolding over hours or days, not minutes. [1] That is the reason the workspace matters. Ona's technology provides secure, persistent environments where agents can access tools, systems, and context over time, while organizations control where the agents run, what they can access, how credentials are scoped, how activity is logged, and how work moves through review. [1]

That description is less glamorous than a demo and more important than one. If an agent works for days, the hard question is no longer whether it can complete a task in a sandbox. The question is what repository it can touch, which secrets it can read, which services it can call, who sees the diff, and what happens when the work crosses from suggestion into deployment.

The procurement layer is already visible. OpenAI says eligible Oracle Universal Credits will be usable for OpenAI models and Codex through OCI in coming weeks, giving Oracle customers a path through existing purchasing workflows and cloud commitments. [2] That makes agent adoption look less like a consumer subscription and more like infrastructure sold through the same channels companies already use for databases, compute, identity, and spend controls.

Visa's agent-commerce page supplies the parallel outside code. It describes AI-initiated transactions that require embedded credentials, controls, authentication, spending limits, approval workflows, and trusted identity signals. [3] The same structure follows Codex into enterprise software: autonomy without authorization is not a product. It is a risk report waiting to happen.

MSM can summarize Ona as another startup acquisition. X can celebrate agents that keep working after the laptop closes. Both frames skip the control plane. The acquisition is interesting because OpenAI is buying the boring part that makes delegation governable: persistent workspaces, reproducible context, customer boundaries, logs, and review points.

The unanswered questions now matter more: price, closing conditions, regulatory review, data residency, credential scope, customer audit rights, and liability when sustained agent work changes production code. Codex does not merely need to be clever. It needs a place to work where the customer can say yes, no, later, never, and prove which answer was given.

-- DAVID CHEN, Beijing