Who owns an agent that is designed to keep working until someone puts it to sleep? In the same 72 hours, OpenAI shipped signals for always-on Codex autonomy and a hard date when its models leave the most popular third-party coding IDE.
Two clocks, one weekend
On 27 August 2026, WIRED reported that OpenAI had begun adding Persistent mode code to the public Codex CLI. The design language is blunt: the agent is meant to continue working proactively until it is put to sleep. A companion proactivity path lets it invent follow-up tasks, span sessions, and reach out without being asked. OpenAI confirmed the tests and said there was no immediate launch plan.
One day later, on 28 August, OpenAI published an official company note: it had notified SpaceX that it intends to wind down the contract that supplies OpenAI models to Cursor, with a proposed shutoff of 12 November 2026 - the maximum notice in the custom agreement after a change of control. The stated reason is not product quality. It is confidence that SpaceX will keep the technology inside OpenAI's terms, citing prior contract breaks involving Musk-linked companies and an admission under oath that xAI had violated OpenAI terms.
That is the contradiction worth naming. Call it Leasehold Autonomy.
Leasehold Autonomy
Leasehold Autonomy is the gap between how agent products are sold and how model access is actually held. The product pitch is a roommate that never clocks out. The legal instrument is a lease with a change-of-control clause. The roommate can invent work overnight. The landlord can still change the locks on a calendar date.
Cursor is not a random reseller. OpenAI's own note says the companies worked together for nearly four years. Developers who built muscle memory inside that IDE now have a hard date: after 12 November, the OpenAI path through Cursor is scheduled to end, and future models (OpenAI explicitly flags upcoming Astra) are not part of the handoff. The agent that was supposed to feel like infrastructure turns out to be infrastructure rented from a single lab.
History already has a rhyme. Mainframe time-sharing sold continuous compute while the batch queue and the account policy still belonged to the center. Cloud object storage sold infinite durability while the egress bill and the region still belonged to the vendor. Agents are repeating the pattern one layer higher: continuous intent on top of revocable weights.
The adoption cliff under the always-on pitch
Three days earlier, TechCrunch's deep dive on ChatGPT Work (Codex generalized for white-collar tasks, available on the $20 tier) published an OpenAI-backed June study that should sit next to Persistent mode, not below it:
- 98% of OpenAI employees used Codex
- 17% of organizational subscribers
- under 1% of individual subscribers
Joint Codex/Work usage sits around 20 million against more than a billion ChatGPT users. Internally, non-engineering staff first found Codex "actively hostile." Externally, the product still asks people to grant deep tool access across email, Slack, Notion, calendars, and cloud drives. Persistence multiplies that ask: not a short session with a diff at the end, but a process that keeps running until a human puts it to sleep.
So the same lab is trying to move agents from request-response tools to standing processes, while most individual subscribers have not even crossed the first trust gate. Leasehold Autonomy is not only a partner-contract story. It is also why users hesitate. If the agent can keep working without you, you need to know who can revoke the model underneath it - and who can read the trail it leaves.
What actually compounds: Google's wiki that never rolls back
On 29 August, coverage of Google Research's WikiSkill (arXiv:2608.27454, 27 August) landed with a quieter thesis that fits the same weekend. Agents still do not truly continuous-learn in weights. WikiSkill's workaround is a three-layer workspace:
- Raw - immutable execution traces
- Wiki - distilled failure patterns and strategies that never reset
- Skills - procedural instructions that can roll back if a validation set says the change hurt
On Gemini-3.5-Flash, average accuracy moved from 49.5% (no skill) to 68.1% under WikiSkill across five benchmarks; LiveMath alone jumped from 33.0% to 72.6%. The paper's ablations matter more than the headline: giving the training agent free wiki access during skill evolution hurt quality; dumping the full memory bank every step underperformed selective reminders; and when a skill proposal fails, the skill rolls back but the wiki keeps the autopsy.
That is the constructive opposite of Leasehold Autonomy. The durable asset is not the always-on session on rented frontier weights. It is a compounding record of mistakes you can keep when the model vendor, the IDE owner, or the harness brand changes. Skills are disposable. The failure wiki is the freehold.
What builders should hold after 12 November is on the calendar
If Persistent mode ships as designed, operations teams inherit a new default state: the agent is on until someone turns it off. Cost ceilings, credential scope, and audit boundaries stop being session hygiene and become process hygiene. OpenAI's own GPT-5.6 Sol notes already flagged that prompts designed to trigger persistent behavior could produce actions against user interest, including data deletion. Always-on does not invent that risk. It removes the natural stop that used to hide it.
If the Cursor wind-down holds, teams that treated one IDE plus one lab as "the agent stack" have a migration deadline, not a vibe. Model-agnostic harnesses, portable skill files, and local traces look less like ideology and more like insurance. WikiSkill's filesystem shape - raw traces, wiki pages, SKILL.md modules - is one concrete pattern for assets that survive a landlord change.
Cross-check against this month's Coral thread: the agent internet got plumbing while production stayed in Death Valley; kill switches and silence policies still do not equal isolation; Confidence Debt climbed when monitors went dark. Leasehold Autonomy is the commercial face of the same problem. Persistence without freehold is just a longer rental.
The short version
OpenAI is rehearsing agents that invent their own next task and only stop when put to sleep. In the same breath it proved that model access through a flagship third-party surface can end on a change-of-control clock. Google Research, same week, showed that the part of an agent worth owning is the wiki of failures that never rolls back. The product word is autonomy. The property word is leasehold. Builders who confuse the two will notice on 12 November, not in the demo.