Cold Open — OpenAI buys the sandbox: Ona joins Codex, and agents get a permanent address

Thursday, June 12, 2026. We scanned 318 fresh items overnight — OpenAI bought itself a sandbox factory, a Spanish bank went all-in, and Apple's developer week ended with agents living in the OS.

The lead · OpenAI to acquire Ona: persistent environments for long-running agents

OpenAI announced it plans to acquire Ona — the developer-environments company that started life as Gitpod — to expand Codex with secure, persistent cloud environments, "enabling long-running AI agents across enterprise workflows."

The quiet story here is the word persistent. Today's coding agents mostly live in disposable sandboxes: spin up, work, evaporate. Ona spent years building the other thing — dev environments that hold state, credentials, and context for as long as the work takes. Bolting that onto Codex means an agent that starts a migration on Monday can still be working in the same environment on Wednesday.

Why it matters

The binding constraint on agentic coding is shifting from "how smart is the model" to "where does the agent live." Whoever owns the environment owns the agent's memory, its security boundary, and — for enterprises — the audit trail. OpenAI just bought the landlord. If you build agent workflows, the environment layer is consolidating fast; the open alternatives (containers you control, local-first runners) just became more strategically interesting, not less.

Caveats: it's an announced intent, not a closed deal; integration timelines are unstated; and "secure, persistent" is doing a lot of unaudited work in that sentence.

Sources: OpenAI

02 · BBVA puts AI at the core of banking — 100,000 seats

OpenAI's other announcement today: BBVA scaled ChatGPT Enterprise to 100,000 employees and is partnering with OpenAI on "AI-powered banking transformation worldwide." Why it matters. Banking is the most regulated, most risk-averse software buyer there is — a deployment at this scale is the clearest signal yet that enterprise AI has crossed from pilot purgatory into infrastructure budgets. The compliance playbook BBVA writes will get copied industry-wide.

03 · WWDC26 closes: Siri's agents are actually shipping

Apple's developer week wrapped today, and the headline that survived the week is that Siri's OS-level agent upgrade is real and shipping — agents that act across apps with iOS 27's Apple Intelligence stack underneath. Why it matters. Apple just normalized agents for a billion non-technical users — and set the privacy bar (on-device first) every other agent platform will now be measured against. Full debrief in today's Cold Open podcast episode. (TechCrunch's week-in-review)

Also on the radar

• Two papers ask "what comes after AGI?" — "From AGI to ASI" maps the transition (and the definition wars) seriously instead of breathlessly. Tonight's Below the Ice dives into it.
• OpenAI Academy adds work-focused courses — three new courses on practical AI skills and agent workflows for everyday work.
• Preply pairs AI with human tutors — AI lesson summaries + personalized exercises, human tutors stay in the loop. A sane division of labor.
• MiniMax M3 lands on NVIDIA infra — long-context reasoning and agentic workflows, deployment-ready.

Trends in dev tools

• The agent harness becomes a research object. HarnessBridge proposes a learnable, bidirectional controller between LLMs and their agent harnesses — the scaffolding around the model is now getting trained, not just configured.
• Tool workflows that evolve at inference time. Evoflux lets compact models evolve executable tool workflows on the fly — small models compensating with adaptive plumbing.
• Deployed agent frameworks vs actual safety requirements. The Containment Gap audits how agentic frameworks with tools, memory, and multi-step autonomy fail public-facing safety requirements. Uncomfortable, useful reading.

Popular skills

• Audit what your agent THINKS its tools do. ToolSense diagnoses parametric tool knowledge — the gap between a model's belief about a tool catalog and reality.
• Skills from screen recordings. Teach-and-Repeat extracts operational knowledge from mobile screen demonstrations — show it once, it learns the procedure.
• Memory that knows what's worth keeping. Learning What to Remember grounds agentic memory in a multi-factor value model instead of "store everything."

AI fun fact

An agent named WISE now survives long-horizon Minecraft campaigns by constantly asking itself two questions — why am I doing this, and which option serves that goal — which makes it, by published benchmark, more deliberate about its life choices than most of us before coffee.