Cursor

Cursor is an AI software development platform made by Anysphere, Inc., combining a code editor with models and agents that search repositories, change files and run development tools. The product brings these pieces together: the interface shapes how developers delegate work, retrieval supplies relevant code, and execution environments let agents test what they produce. By September 15, 2026, its scope included interactive editing, cloud agents and a beta rollout of coordinated, recurring Projects. Anysphere is the legal maker identified in Cursor's terms; corporate history and ownership announcements provide context for the product.[1][6][12]

The programming environment grew out of earlier experiments rather than a fixed original blueprint. In a Latent Space interview, Aman Sanger recalled working with Michael Truell on a CAD-assistance project also called Cursor. In his account, training data was scarce, models overfit as they grew, spatial reasoning proved difficult, and CAD integrations imposed their own obstacles. These are recollections of that team's experiments, rather than general limits on AI for CAD. The account helps explain the team's eventual move toward software, where code and development tools offered a different setting for AI assistance.[3]

In the founding-team interview with Lex Fridman, Truell recalled that early access to GPT-4 near the end of 2022 widened the team's ambitions: instead of a collection of specialized tools, they could pursue a broader programming environment. The team chose a VS Code-derived editor because owning the editor gave them control over interactions that an extension could constrain. That decision made the surrounding workflow part of the product, alongside the model answering a prompt. The interview records the team's explanation; its separate hosting does not independently corroborate that account.[4]

The team also treated latency as a design constraint. Arvid Lunnemark's phrase “Fast is fun” captures a practical test: an experiment had to feel enjoyable to use. In an editor, waiting changes whether a developer keeps exploring an idea or abandons the interaction. Cursor's later model work likewise emphasized efficient tool use and shorter completion times as training objectives.[4][6]

The original Composer model, announced on October 29, 2025, followed an experiment called Cheetah that explored faster agents. Cursor's technical account described Composer as a mixture-of-experts model specialized through reinforcement learning in coding environments. Its training tools included file editing, terminal commands and semantic search; rewards encouraged efficient tool choices and parallel work. In this training process, the model attempts coding tasks, and rewards for those attempts guide updates that make rewarded behavior—such as using tools efficiently—more likely in future attempts. The account also describes asynchronous reinforcement learning using PyTorch and Ray. The engineering connection is consequential: the team adapted Cursor's Background Agents VM scheduler for training environments, bringing infrastructure built to run agents into the process of teaching a model to use those environments.[6]ⁿ¹

Finding the right code is another part of that workflow. A query about authentication handling, for example, may have no literal match in a repository. Cursor's semantic-search account, published on November 6, 2025, describes custom embeddings—numerical representations that help match a query with related code—and indexing to retrieve code by meaning, complementing regex search. Its described training method starts from agent traces: identify content that would have helped earlier, have a language model rank that content, then train an embedding model to learn the resulting similarity judgments. This makes retrieval a component that can improve from the agent's own work.[7]ⁿ²

The March 27, 2026 Composer 2 technical overview identifies Kimi K2.5 as its open base model, followed by continued pretraining emphasizing code and reinforcement learning with Cursor's deployed tools and harness—the tools and execution setup surrounding the model. That provenance distinguishes the base model from the team's subsequent specialization. The overview also credits Fireworks and Colfax collaboration. Across these accounts, the model strategy ties training to the environment in which users ask agents to work, including terse requests and changes spanning multiple files. These details are supported by the inspected overview; the full linked report remains uninspected.[10]

In his February 26, 2026 essay “The third era of AI software development”, Truell proposed a progression from autocomplete, through synchronous agents, to independent cloud agents. This is his product thesis, rather than an established industry periodization. Its engineering implication is that, as work runs for longer without continuous supervision, developers spend more effort defining the problem and setting review criteria. Tests and environments become especially important. A broken setup that a human can work around may instead cause repeated failures when agents operate at scale.[8]

Josh Ma's June 2, 2026 cloud-agent engineering account makes that implication concrete: missing dependencies or access needed for verification can quietly degrade an agent's output. The account describes infrastructure for checkpointing virtual machines (VMs), saving an execution environment so work can resume, and for restoring, forking and hibernating them while controlling networking and credentials. Execution also needed to survive interruptions. Ma recounts that an early work-stealing architecture proved fragile, and the team moved to Temporal, a system for preserving workflow progress and resuming execution after failures, rather than rebuilding those durable execution primitives. Reliable unattended coding depends on preserving and resuming the work as well as generating the next action; the operational experience described here comes from the product's own engineering team.[9]

Self-Hosted Machines, announced on September 2, 2026, lets tool execution run on customer-managed machines near private services or specialized hardware. A worker connects outward over HTTPS, receives tool calls, executes them and returns results. Only the execution environment moves: inference, planning and the agent loop remain in Cursor's cloud. Returned tool output may contain code, and Cursor may process or store transcripts. The feature therefore gives customers control over where tools run while retaining cloud model processing.[11]

Projects, which began its beta rollout on September 10, 2026, extends this approach to coordinated and recurring work. A coordinator delegates coding to subagents, shared context files synchronize across cloud and local machines, and subscriptions trigger scheduled or event-driven tasks. Cursor's internal design-system example describes an engineer first correcting proposed fixes, then having the coordinator monitor new pull requests and add lint rules for repeated mistakes. The useful mechanism is a feedback loop: corrections can become shared instructions and automated checks that guide later work.[12]ⁿ³

Cursor's terms updated September 3, 2026 identify Anysphere, Inc. as its maker. The Lex Fridman interview introduces Michael Truell, Sualeh Asif, Arvid Lunnemark and Aman Sanger as historical founding members of the company behind Cursor. This is company background, rather than a corporate founding claim about the software product or a current leadership roster. Lunnemark announced his decision to leave Cursor on October 17, 2025; his personal announcement does not establish an exact last working day or subsequent role. On August 14, 2026, the Cursor team announced that SpaceX had completed its acquisition, distinguishing the company transaction from an earlier April partnership with SpaceXAI. The team framed the acquisition as a way to expand model-training compute and reduce inference costs. Those are attributed strategic aims; the announcement does not establish their eventual effect on product users, and transaction price and detailed terms remain unsupported by the inspected evidence.[1][4][5][2]

Following the company acquisition, OpenAI announced on August 28 its intention to wind down its direct model-supply agreement with Cursor, saying future models would not be provided and proposing November 12, 2026 as the shutoff date. As of September 15, that date was prospective and unconfirmed: OpenAI's transition guidance said official termination awaited confirmation between the companies and Cursor could end direct access sooner. The guidance distinguished alternative local access inside the editor from model supply for Cursor-managed services. That distinction matters for the product's available workflows, alongside its Composer models and the infrastructure built around them. Both transition sources originate with OpenAI and do not independently corroborate each other.[13][14][10]ⁿ⁴