Aman Sanger and the Art of Shipping the Unfinished

At fourteen, Aman Sanger began writing code. A decade later, he was helping to change where code gets written, how it appears, and whether a human needs to write much of it at all. The distance between those two moments looks short only on a résumé. In reality it contains a New York childhood, squash courts, pure Vim, a hedge-fund internship, an AI consultancy, a failed experiment in mechanical design, and a very consequential decision to stop waiting for the old code editor to catch up.

Sanger is one of the four MIT friends who founded Anysphere in 2022. Its product, Cursor, began as a wager: language models would become so useful at programming that sprinkling autocomplete onto an existing workflow would not be enough. The whole room needed rearranging. Cursor, built from the open-source foundation of VS Code, put conversation, code generation, editing, retrieval and eventually autonomous agents inside the place where programmers already worked.

The company’s ascent has since acquired the usual scenery of a Silicon Valley fable: famous investors, vertiginous valuations, enterprise customers and a 2026 deal that folded Cursor into SpaceX. Yet the more instructive Sanger story is not the financial one. It is about how often serious work begins as an experiment that fails its audition.

The useful wrong turn

Before Cursor, the founders tried teaching models to assist with computer-aided design. Imagine Copilot inside software such as SolidWorks, anticipating the next piece of a three-dimensional object. The concept was neat. The data was not. There simply was not enough good material on which to train the system, and the result never became the product they hoped it might.

They moved toward code, where the evidence was abundant and the early models were improving quickly. GitHub Copilot had already delivered a small jolt. Sanger, a devoted user of pure Vim and a terminal, switched to VS Code just to use it. That conversion mattered. A tool did not have to respect an engineer’s long-established ritual if it offered enough delight to replace it.

A lot of the work for Cursor has been just experimenting with what is possible.Aman Sanger, in conversation with Peak XV

When the team received early access to GPT-4 near the end of 2022, the larger possibility sharpened. These systems would not remain a clever completion box. Programming itself would pass through them. So the founders chose the expensive, awkward route: an entire editor rather than another plug-in. In January 2023, Sanger introduced Cursor publicly as an IDE built from the ground up around language models. The first pitch was direct. Tell the editor what to do and watch.

Fast enough to feel like play

Sanger’s public work shows a person preoccupied with the unglamorous physics of making AI useful. How much context can a model hold? How quickly can it generate an edit? What is the memory cost of attention? Can a prompt be compiled against a strict token budget? These are not cocktail-party questions. They are what stand between an impressive demonstration and a product someone opens on Monday morning.

His 2024 technical essay on editing files at 1,000 tokens per second split the task into planning and applying. A powerful model could decide what the change should be; a specialized model could execute the change quickly. The distinction sounds mechanical, but its purpose was emotional. Delay breaks concentration. Speed preserves the little current of pleasure that makes building things on a computer addictive.

This also explains the team’s appetite for throwing work away. A feature that is technically possible may still be slow, confusing or joyless. One surviving interaction can sit on a small graveyard of rejected versions. The method resembles good editing: remove the sentence that proves you are clever if it spoils the paragraph.

The virtue of being nearly right

At MIT, Sanger made a bet with Shengtong Zhang, a gifted mathematician in the founders’ orbit. By the summer of 2024, he thought, a model would win a gold medal at the International Mathematical Olympiad. His friends considered this fanciful. The relevant system eventually landed one point below the gold threshold.

“Technically incorrect but one point away,” Sanger said later, a perfect founder’s verdict: the letter of the bet conceded, the curve vindicated. Michael Truell remembered him wearing a scaling-laws T-shirt decorated with formulas and charts. One can imagine the garment doing little for campus fashion and quite a lot for the argument.

Sanger’s optimism was never merely that models would become larger. He was interested in domains where answers could be checked. Mathematics has proofs; software has tests, compilers, interfaces and users who can say, often with impressive speed, that something is broken. Verification gives learning a rail to run on. It also makes programming unusually fertile ground for capable agents.

The future of review is not code. It’s agents proving their work.Aman Sanger, February 2026

That line describes a deeper change than better autocomplete. In Sanger’s view, an agent should not merely produce a diff and leave a human to inspect the debris. It should operate the thing it built and return with evidence: a video of the feature working, perhaps, or another concrete demonstration. The engineer moves upward from keystrokes toward intent, delegation and judgment.

A bookshelf with sharp elbows

Founders tend to publish principles. Sanger publishes a reading log, which is much more dangerous. It is heavy on science fiction, computing history, biography and ambition: Dune, The Dream Machine, Snow Crash, The Wright Brothers, Robert Caro. Almost every title receives five stars, making the tiny comments do the real critical work.

The list punctures the familiar portrait of the bloodless technical founder. Sanger can be impatient, sentimental, amused and blunt within four entries. After reading Atlas Shrugged, he recorded not an ideological awakening but a scheduling consequence: the next month brought 100-hour weeks. His entry on a Murakami running memoir confesses that competitiveness got in the way of recognition. The jokes are dry because the feelings are not.

That competitive instinct appears elsewhere. He played squash seriously at MIT. He watches rivals closely and has said they can expose ideas Cursor missed. This is competition without the theatrical contempt common to technology. The opponent is useful. It raises the standard and occasionally lends you a feature worth reconsidering.

What an engineer becomes

By late 2025, Sanger was talking less about whether AI could code and more about what organizations would do with the abundance. The unimaginative answer is fewer junior hires and the same backlog. His preferred answer is more ambitious projects. If software gets cheaper to create, companies can attempt systems they previously could not justify. Productivity becomes permission rather than a euphemism for subtraction.

Cursor’s own expansion followed that pattern. It grew from completing lines to editing across files, answering questions about whole repositories, running tasks in the background and training specialized coding models. Sanger moved with it from inference details into research leadership and public arguments about agents. In 2026, after SpaceX exercised an option to acquire Cursor, his public biography changed to “Building @SpaceXAI | @cursor_ai founder.” The original workshop had acquired a much larger roof.

There are tensions in this future. A tool that makes production effortless can generate mountains of undistinguished software. An agent that works independently can be independently wrong. Sanger’s answer is not blind trust. It is a better loop: richer context, faster correction, visible proof and a human who still knows what good looks like.

That last requirement may be the durable one. When typing becomes scarce, taste does not. The engineer of tomorrow might write fewer functions, but must frame better problems, detect weak answers and decide what deserves to ship. It is rather like Sanger’s bookshelf: the stars are easy. The one-line judgment is the difficult bit.

His career so far is a sequence of doors opened before the room behind them was finished. Start coding. Try a company. Test CAD. Abandon it. Build the editor. Train the model. Ask the agent to show its work. The unfinished state is not an embarrassment in this method. It is where information enters.

Sanger’s wager is that software will become easier to make and harder to judge. If he is right, the best programmers will resemble editors, conductors and exacting customers of machine intelligence. If he is technically wrong but one point away, that may once again be close enough to matter.