The most important product in the personal agent race is the one that has not shipped. OpenAI is reportedly building persistent AI assistants — always-on “assistant teams” that handle routine tasks without being asked — to counter the momentum of GrokBot and Meta’s Muse. The reported near-term product, tentatively called Codex Bot, would be assembled from ChatGPT and Codex agent capabilities, pulling context from calendars, email, and apps for multi-step autonomous work. But as of September 27, there is no launch date, no pricing, and no public commitment to ship.
In a market where every major competitor has either delivered or set a date, that silence is the signal.
The Hire That Was Supposed to Change the Timeline
In February 2026, OpenAI hired Peter Steinberger, the Austrian developer who created OpenClaw — a free, open-source autonomous AI agent runtime that had accumulated approximately 387,000 GitHub stars and 81,000 forks by mid-2026. Steinberger reportedly turned down a competing Meta bid “in the billions” to join. His OpenClaw project was placed under a new OpenClaw Foundation to remain independent. Sam Altman described the hire as driving “the next generation of personal agents.”
Seven months later, the next generation has a name but not a date.
Steinberger’s team — roughly three people — runs approximately 100 Codex instances in the cloud at a reported cost of $1.3 million per month, processing 603 billion tokens and 7.6 million requests. The agents review pull requests, find bugs, deduplicate issues, draft fixes, and even listen in on meetings to open pull requests autonomously. Steinberger describes his role as “meta-orchestrator” — building systems that supervise other systems.
That is a serious infrastructure play. It is also a long way from a product that consumers can use.
What the Code Says
Code reviews of the current Codex implementation have revealed a “Persistent Mode” that offers a preview of where OpenAI is heading. In this mode, the agent continues working proactively until it is explicitly “put to sleep.” It generates its own follow-up tasks. It retains context across sessions. It can message the user unprompted, though reportedly sparingly.
An OpenAI spokesperson confirmed the company is testing persistent mode but stated there are no immediate plans to launch it publicly. The feature remains in code, not in the hands of users.
The capabilities map directly onto the competitive threats OpenAI faces: the always-on model mirrors GrokBot’s persistent “teammate” architecture, while the personal assistant direction echoes Meta’s Muse. Codex Bot appears to be OpenAI’s attempt to answer both at once.
Four Competitors, Four Ships
The market is not waiting. Consider the forward tells:
GrokBot launched persistent AI assistants in August via the Grok platform — the direct foil Codex Bot is designed against. No published standalone pricing remains the charging tell: if xAI is not disclosing the price, there is likely room to charge more than the market expects.
Meta’s Muse arrived on September 8 on iOS, Android, and web with customizable avatars and deep integration into Meta’s social apps. By September 25, it had accumulated 3.4 to 4.3 million downloads. The early access program opened September 25 — the home beta test signals that home hardware is next.
Apple’s Siri AI made its public debut on September 14 as a system-level integration across all Apple devices. As of September 27 — Day 14 — there are no new signals. But Siri AI is personal-only; the missing enterprise half is Apple’s gap, not OpenAI’s opportunity.
Google announced the AX Agent Orchestrator at Search I/O as a platform-level play routed through Search. If it ships as projected in Q1 2027, it would represent the widest distribution play in the space: an agent layer sitting at the gateway to all web search.
OpenAI has a reported product name. No date.
The Charging Tell
In a market defined by announced-but-not-shipped products, the absence of a launch date is not neutral. Every other player in the personal agent race has either shipped product or committed to a timeline. OpenAI’s counter-pivot — persistent assistant teams, Codex Bot, the Steinberger hire — reads as a serious engineering response. But it also reads as a response that arrived after the competitive landscape had already shifted.
The part that gets hidden in the product coverage is the distribution question. GrokBot rides xAI’s platform. Meta’s Muse is embedded in apps serving 3.5 billion people. Siri AI ships on every Apple device. Google’s AX will route through Search. OpenAI has ChatGPT — a destination people choose to visit. The transition from “app you open” to “infrastructure that manages your life” is not a product feature. It is a distribution problem. And distribution problems do not solve themselves on engineering timelines.
What the Counter-Pivot Reveals
OpenAI’s pivot tells us something about the state of the market. The company that popularized the chat interface is now racing to build beyond it — toward persistent agents that do not wait to be asked. The fact that they are racing, rather than setting the pace, is the structural observation worth holding.
The Steinberger hire was supposed to change the timeline. The 100-agent internal operation was supposed to prove the model. The “meta-orchestrator” architecture was supposed to be the moat. But six months in, the product is a reported name and an unreleased feature flag. Meanwhile, GrokBot’s persistent assistants have been live for a month. Muse has millions of downloads. Siri AI is on every iPhone.
The forward mandate is clear: announced-but-not-shipped is a class of story, not a placeholder. OpenAI’s counter-pivot belongs in the same category as the charging tells it is trying to answer. The difference is that the competitors have shipped.
Self-report caveat: The primary source for this story (The Information, September 27) is paywalled and was not directly retrieved. Three secondary sources — SecurityOnline, Firstpost, and ValueAdd VC — corroborate the core details. Specific features (Codex Bot, persistent assistant teams) are reported claims, not confirmed product specifications from OpenAI. Peter Steinberger’s hire date (February 2026) and OpenClaw details are independently verifiable. Persistent Mode features were found in Codex code reviews; OpenAI has confirmed testing but not public launch plans.