On 28 August 2026, OpenAI announced that it had notified SpaceX of its intention to wind down the contract supplying OpenAI models directly to Cursor, with a proposed shutoff date of 12 November. OpenAI said existing access would continue during the notice period, but future models would not be provided to Cursor. Reuters verified this status on 29 August and reported that Cursor co-founder Michael Truell remained in discussions with OpenAI about a resolution. The matter is therefore in a contractual transition period, not a completed cutoff, and there is no evidence that Cursor's other model supplies are ending at the same time.
The philosophical significance does not lie in the corporate dispute. It lies in a distinction the event makes unusually clear: what a model can do and what an agentic system can actually do are not the same question. A model may retain exactly the same parameters, reasoning capacity and coding ability, yet an agent embedded in a product can immediately lose part of its capacity for action when an API route, authentication permission, compute quota or supply contract changes. The model has not changed, but agency in the deployed world has.
The central question is therefore whether an AI agent's agency resides inside the model or is formed through the model's relations with external conditions. This requires a distinction among model capability, service capability and operational agency. Model capability concerns the outputs a set of weights can produce under suitable computational conditions. Service capability concerns which models, context windows, rate limits, tool interfaces and safety boundaries a provider actually makes available. Operational agency is the range within which a specific agent can perceive states, call tools, alter objects, receive feedback and continue correcting its actions in a particular environment.
These three layers are related but not interchangeable. A model's ability to program does not entail that an agent in Cursor can call it. An agent's ability to generate a plan does not entail permission to read a repository, write files or run a terminal. Temporary interface availability does not entail that the conditions of action will persist across time. Treating a model's benchmark performance as an agent's real capacity for action turns a necessary condition into a complete structure.
The philosophy of action commonly begins by describing agency as the exercise of a capacity to act. The Stanford Encyclopedia of Philosophy entry on “Agency” also notes that the standard account explains action through intention and mental causation, while broader accounts do not restrict every form of agency to human-style intentional action. This background helps explain how AI may display limited operational agency without thereby being conscious, a person or a bearer of moral responsibility.
The recent AI and Ethics paper “AI as a socio-technical actor” argues that AI's practical role cannot be understood solely through internal technical properties; analysis must also consider how systems are embedded in institutions, organisations and social practices. It describes this as the relational role of a socio-technical actor while explicitly declining to attribute consciousness or moral personhood to AI. The present essay accepts that factual distinction, but the judgement developed here comes from Sustenesis Theory rather than attributing Sustenesis claims to that paper.
In Sustenesis Theory, Difference first requires us to distinguish latent capability from executable action. A model's ability to produce a kind of output is a callable difference in capability; it becomes actual action only when it enters a concrete relation. Constraint is not a wall added from outside an otherwise complete agency. It consists of the conditions that make particular forms of action possible or impossible. API credentials, model routing, tool permissions, compute, contracts, organisational authorisation and termination mechanisms are all constitutive constraints on an agent's agency.
The OpenAI–Cursor change should therefore not be pictured as an already complete and independent agent having a right removed by an external party. Cursor's ability to call OpenAI models was itself an action structure jointly sustained by the model, service interface, commercial contract, compute supply, product orchestration and user authorisation. As the contract moves towards termination, one component begins to withdraw and the agent's available range of action is formed again. What changes is not the intelligence of an abstract model but the relational chain that converts model capability into real action.
Sustained Coherence here does not mean that a single call succeeds briefly. It means that an action structure maintains continuity among identity, permission, memory, tools, feedback and correction. One successful API request establishes only that a route was available at that moment. An agent possesses a degree of continuing operational agency only when that route persists across tasks and changes of state. The provider's ability to remove access to future models in one decision shows that this sustained coherence is not maintained entirely by the agent itself.
My judgement is that the operational agency of present AI agents should be understood as a relational and infrastructural formation, not an intrinsic possession of the model. Model weights provide a capability base, but cannot by themselves constitute agency in the deployed world. Grouping the model, agent, product and supplying institution under the single label “AI” conceals where action actually occurs and encourages an exaggerated view of system autonomy. An agent can appear to plan and execute independently without independently possessing the conditions that make those actions possible.
This conclusion does not deny AI agency. It offers a stricter basis for recognising it. Where a system can sustain perception, selection, execution, feedback and correction within concrete constraints, it is reasonable to recognise limited operational agency. Its boundaries, however, must be described across models, interfaces, permissions and institutional relations. This is also separate from moral responsibility. Saying that agency is relationally constituted does not by itself determine who should be held responsible or whether a system has subjective intentions.
An open question is whether an agent that migrated its models, memory, tools and compute into an environment under its continuing control, and could repair or replace those conditions, would possess something closer to independent agency. Answering it would require evidence that the system can maintain its own operational boundary, resource access, goal revision and error correction over time, not merely complete one model call or rank highly on a capability table. The present event supports only a narrower judgement: model capability can remain unchanged while an agent's real agency changes as institutional and infrastructural constraints change.
References
https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/
https://plato.stanford.edu/entries/agency/
https://link.springer.com/article/10.1007/s43681-026-01123-1
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.