Ben Okonkwo: Mm. Two moves, one week. I'm still — honestly, I'm still working out which one is the bigger story.
Jonathan Ingles: They're the same story. That's the point. June 11th, OpenAI buys Ona — formerly Gitpod, 60 people in Berlin, reportedly hundreds of millions of dollars. Sarah Friar breaks it on LinkedIn. Then seven days later, June 18th, they ship Record & Replay inside Codex version 26.616 — show the agent a task once, it learns it, no code. And they block it from the entire European Economic Area, the UK, Switzerland. From launch.
Ben Okonkwo: Wait — from launch? Not a rollout delay?
Jonathan Ingles: Day one exclusion. And Sam Altman is out there calling Ona 'the hands that get things done' — giving every Codex agent 'a full runtime.' Persistent cloud execution, hours or days. That's the pitch.
Ben Okonkwo: Now, they're also claiming five million weekly Codex users and 400% usage growth into the acquisition announcement. That figure — right, that's the one I want to pull apart. Because 400% of what, exactly?
Ben Okonkwo: Right, and that's before we even get to what Record & Replay actually is versus what Automations already does. Because those are two different things, and the coverage is mashing them together. Automations — that's already in Codex. Scheduling recurring tasks, running them in isolated cloud sandboxes connected to GitHub. That's existing. What Record & Replay adds is... it's a different paradigm. You perform a task once, the agent watches, and it converts that observation into a reusable skill. No prompting, no workflow config. Learning from demonstration.
Jonathan Ingles: But what does 'editable skill' mean when it breaks?
Ben Okonkwo: That's — yeah, that's exactly the gap. The sources don't say. No failure rate, no edge-case data. Are users debugging Codex's interpretation, or are they basically writing the code they thought they were avoiding? Silence on that.
Jonathan Ingles: And that's where Ona's actually new. Ona isn't fixing Record & Replay's failure modes — it's solving duration. Sarah Friar said it plainly: hours or days, not minutes. That's the Gitpod infrastructure. That's the genuinely new thing.
Ben Okonkwo: Persistent cloud execution. An agent keeps running after you close your laptop. That is a qualitative shift for enterprise deployment — I'll grant that. The orchestration problem it solves is real.
Ben Okonkwo: But back to 400%. Active users? API calls? Task completions? The number is compatible with millions of people opening Codex once and never returning. That's not proof-of-value — that's a denominator problem.
Jonathan Ingles: And that denominator problem is exactly what makes it a justification metric. OpenAI needs to explain a deal reportedly worth hundreds of millions of dollars. What do you reach for? The biggest number you have. Four hundred percent. It doesn't survive scrutiny — it's not designed to.
Ben Okonkwo: Right, but — I mean, noisy isn't the same as meaningless. If the baseline and methodology are eventually disclosed, even a rough signal tells you something about trajectory. The stronger claim, that it's deliberately fabricated to justify the price, that requires intent I can't read from the number alone.
Jonathan Ingles: No, I don't buy that.
Ben Okonkwo: Why not?
Jonathan Ingles: Because the EEA exclusion is the tell. Record & Replay is blocked from launch — day one — in the exact jurisdictions with the enforcement infrastructure to actually demand that methodology. You don't get to claim the 400% proves enterprise value while simultaneously hiding the feature from every regulator positioned to audit it. That's the contradiction. That's the arbitrage.
Ben Okonkwo: Okay, that — actually, the exclusion is real, the pattern is real. But motive is doing a lot of work there. Compliance caution is a legitimate reading. 'We don't know if Record & Replay clears UK data-protection standards yet' — that's not the same as 'we're staging a regulatory play.' The sources don't distinguish those two.
Ben Okonkwo: Right, but — okay, here's what I keep not being able to resolve. The story isn't whether OpenAI eventually brings Record & Replay to the EEA. It will. That's not the question. The question is what happens when it does — because by then, the feature has already been running in enterprise clouds in unregulated markets for however long. A year, maybe more. And whatever guardrails OpenAI retrofits to satisfy regulators at that point... they're being designed after the system is already embedded. That's not accountability architecture. That's compliance theater built on top of infrastructure that's already load-bearing.
Jonathan Ingles: And regulators will be asking that exact question from behind. That's the position Brussels ends up in.
Ben Okonkwo: Which is — I mean, that's the thing neither the Ona acquisition announcement nor the Record & Replay launch actually answers. When compliance arrives a year late to infrastructure that's already running, does it change anything, or does it just get absorbed?
