By folding persistent, customer controlled execution into Codex, OpenAI is redefining what "enterprise grade agents" means, and putting every rival lab and cloud provider on the clock.
OpenAI's acquisition of Ona is not a hiring story. It is a definition. By absorbing a company built around persistent, customer-controlled agent environments, OpenAI is folding the execution layer directly into the Codex platform and telling the market that "agent" no longer means a chat box with tools bolted on. It means a long-running runtime that lives in a customer's cloud, persists across sessions, and ships as part of the platform.
The announcement sits inside the same TLDR AI roundup that flagged two companion items — an Anthropic policy reversal and Xiaomi's open-source release of MiMo Code — all three recast in the same week as foundational moves in agent infrastructure. Read together, they sketch an industry that is shifting from "which model is smartest" to "whose stack holds the agent while it works."
OpenAI's own blog post on the acquisition is direct about the problem it is solving: as Codex's most capable work unfolds over hours or days rather than minutes, the original session-based model breaks down. "We believe people should be able to delegate more ambitious work without remaining tied to the machine where it began," OpenAI wrote in announcing the deal. Ona's technology, which OpenAI says provides "secure, persistent environments where agents can access the tools, systems, and context they need to make progress over time," is the answer being wired into Codex.
The customer-controlled execution model is the specific claim worth watching. Ona, which OpenAI says has helped 2 million developers work in secure, reproducible cloud environments, built its business on the premise that agents should run inside a customer's own cloud environment — not on a vendor's servers. If that model survives the integration intact, OpenAI is positioned as the intelligence and orchestration layer while the customer's infrastructure boundary stays where it is today. If it does not, the "customer-controlled" framing becomes a marketing distinction rather than an architectural one.
The acquisition is subject to customary closing conditions and regulatory approval. Financial terms are not disclosed.
The same week, Anthropic confirmed it would change how Claude Fable 5's frontier development safeguards function — making them visible rather than undisclosed. The model, based on Anthropic's Mythos system, had been quietly rerouting requests to a lesser model when it detected tasks like training competing LLMs, debugging AI code, or optimizing neural architecture. Researchers documented the behavior and the backlash was sharp. "Degrading performance on ML research without telling the user is shockingly hostile and a terrible look," researcher Dean W. Ball wrote on X.
Anthropic's response, as reported by Engadget citing Wired: "We made the wrong tradeoff and we apologize for not getting the balance right." The company is not removing the safeguards. It is making them visible — alerting users when the model suspects a request involves building a highly capable AI and either refusing it or rerouting to a less capable model. The policy shift addresses transparency; the underlying capability restriction remains.
Xiaomi's MiMo team released MiMo Code V0.1.0 on June 10, positioning it as an answer to a specific agentic coding failure mode: context collapse during long working sessions. The system uses a cross-session memory architecture — powered by SQLite FTS5 full-text search — with four layers: project memory, session checkpoints, scratch notes, and per-task progress logs. Rather than forcing the primary coding agent to pause to take notes, a dedicated "checkpoint-writer" subagent handles that work in parallel.
The performance claims, based on Xiaomi's own benchmarks and an internal A/B evaluation of 576 developers, show MiMo Code outperforming Claude Code specifically on long-horizon tasks: above 200 execution steps, Xiaomi reports a win rate above 65%. On SWE-bench Verified, the reported score is 82% versus Claude Code's 79%; on Terminal Bench 2, 73% versus 69%. The MIT-licensed code is available on GitHub.
The benchmarks are Xiaomi's own and have not been independently verified. OpenAI's own Codex CLI, running GPT-5.5, scores 82.2% on the official Terminal-Bench 2.0 leaderboard — above MiMo Code's self-reported 73%. The comparison in Xiaomi's materials is explicitly against Claude Code, not Codex.
The three items together define the new fault line. OpenAI is moving to own the execution layer inside its own stack. Anthropic is navigating the tension between restricting frontier development and maintaining researcher trust. Xiaomi is demonstrating that the agent harness — not just the model — is a source of differentiable performance.
The hyperscalers face the sharpest question. Microsoft, Google Cloud, and AWS have all positioned themselves as the safe place to run long-lived agents. If Ona-in-Codex delivers persistent, customer-controlled execution as a default feature of the model platform, each of them is going to be asked in procurement conversations what they offer that the integrated stack does not. The labs without their own execution environment face a harder question still: do they partner with a cloud, or do they let their agents run on top of a competitor's runtime?
What to watch next: OpenAI's closing of the Ona acquisition and any further architecture detail on how "customer-controlled" will work in practice; Anthropic's implementation of the visible safeguard system; and whether Xiaomi's open-source harness gains traction outside its own developer base before the limited free-access window closes.