Sualeh Asif and the Context Problem Behind Cursor
Sualeh Asif's public Cursor writing makes a core problem of AI coding unusually visible: a model is only useful when the editor can retrieve the right repository context quickly enough to support an interactive workflow.
AI coding quality depends on what the model can see
A language model can be excellent in isolation and still produce poor code if it lacks the relevant types, nearby conventions, test failures or historical context from the repository. Cursor co-founder Sualeh Asif made this problem explicit in an October 2023 post titled “Our problems.”[1]
The post is valuable historically because it documents the engineering questions behind the product rather than only describing end-user features.
Asif described the editor as a dense field of possible context
The 2023 post lists open files, semantically similar code, symbol relationships, linter outputs, execution traces, Git history, typing history and external documentation as possible context sources.[1]
The challenge is that all of this information cannot simply be sent to a model indiscriminately, especially when context windows and latency are limited.
Retrieval is a ranking problem
Asif described gathering a large candidate pool of tokens and using a fast reranker to select the subset most relevant to the user’s current request.[1]
The best context changes from request to request
A refactor may need symbol relationships and tests, while a documentation question may need comments and external docs. Context selection is therefore dynamic, not a one-time repository summary.
The context problem is also an infrastructure problem
Asif explicitly noted that selecting relevant context was becoming an infrastructure challenge as well as a model problem.[1] Search indexes, embeddings, parsers and rerankers must respond quickly enough to preserve interactive editor latency.
This helps explain why successful AI coding products require substantial systems engineering around the underlying foundation model.
Cursor also treated editing as different from next-token completion
The same post asked for a “copilot for edits,” observing that ordinary autocomplete is useful when writing new code but less helpful for small changes to code that already exists.[1]
Edit prediction requires understanding both the current codebase and the user’s evolving intention.
Repository-scale editing changes the unit of prediction
The model may need to modify a distant file, update several references or infer that one edit implies another. This is a different interaction problem from finishing the current line.
Fast application of edits matters to the human experience
If the user must manually copy every suggested change, natural-language direction becomes cumbersome. Product design has to connect model output to precise source modifications.
Cursor’s later growth validated the importance of these hidden systems
The company’s 2024 Series A post highlighted multi-billion-file retrieval systems, next-edit prediction models and fast code rewrites as core technical achievements.[2]
These are the invisible systems that make an apparently simple conversational request feel responsive.
Asif later documented the team building around those problems
In an April 2025 post, Asif highlighted engineers and researchers working on inference, distributed systems, model training and editor design.[3]
The article reinforces the point that an AI coding product is a stack of models, infrastructure and interface work rather than a thin wrapper around an API.
The context problem grew as Cursor moved toward agents
By 2026, Truell and Asif described an interface coordinating local and cloud agents, integrated browsing and deeper repository interaction.[4] The more autonomy an agent receives, the more consequential context selection becomes because the system is taking actions rather than merely suggesting text.
Cursor’s Series B statement about automating code shows the strategic direction into which these context systems were feeding.[5]
Why Sualeh Asif belongs in the vibe-coding lineage
Vibe coding looks effortless at the surface: the user says what should happen and software changes. Asif’s writing exposes the engineering required underneath that illusion.
His contribution to the historical record is to make context retrieval, reranking and edit prediction visible as core enabling technologies for conversational software development.
Works Cited
- 01
- 02Cursor — Series A and Magic (Aug. 22, 2024) cursor.com
- 03
- 04
- 05
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.
Submit a research lead