Shadow Workspaces for AI Code Iteration

12 posts · Oct 2023 – Jul 2026

What this trail covers

Introduced shadow workspaces to enable AI code iteration within a development environment without affecting the user's primary coding experience. This involves creating isolated, hidden instances of the development environment to allow AI agents to receive language server protocol (LSP) feedback and execute code independently. The initial implementation uses a hidden Electron window, with future plans for a kernel-level folder proxy. This post details the design criteria, the current hidden Elec

All Cursor trails

Jul 2026How Vercel used Cursor to build QueuesThis post details how Vercel used Cursor's shadow workspaces and intelligent context management to build Vercel Queues. It highlights the ability of agents to perform end-to-end testing against real AWS environments, including connecting to services, reading side effects, injecting faults, simulating bugs, and verifying results. The post emphasizes the retention of critical details during context compaction, enabling agents to handle complex, long-running tasks. It also showcases the integration of code review and stacked PRs for incremental review, and the use of cloud agents for asynchronous work delegation. The impact on engineering velocity is quantified with metrics like PR throughput increase and cycle time decrease.
Oct 2025Introducing Cursor 2.0 and ComposerIntroduced Composer, a frontier AI model optimized for low-latency agentic coding, completing most turns in under 30 seconds. This model was trained with tools like codebase-wide semantic search for better understanding of large codebases. Also introduced a new multi-agent interface centered around agents rather than files, enabling parallel execution of multiple agents powered by git worktrees or remote machines. Developed native browser tools for testing agent work and iterating until correct results are produced, addressing bottlenecks in code review and testing.
Oct 2025Introducing Plan ModeIntroduced Plan Mode, a new capability for AI agents to create, research, and update plans for complex coding tasks. This feature includes an interactive editor for inline plan modification and integrates codebase research to find relevant files and ask clarifying questions. Plans can be saved as Markdown files within the repository.
Sep 2025Improving Cursor Tab with online RLImplemented online reinforcement learning to improve the Cursor Tab model. This involved defining a reward function that encourages accepted suggestions and penalizes rejected ones, and using the Policy Gradient Theorem to update the model based on user interactions. The system now deploys new model checkpoints frequently (1.5-2 hours) to collect on-policy data for continuous training, resulting in a new default Tab model that makes 21% fewer suggestions with a 28% higher accept rate.
Aug 2025Cursor Agent CLIThe Cursor Agent has been extended to be usable via a Command Line Interface (CLI), enabling headless operation in any environment. This allows for integration with other IDEs and the execution of multiple agents in parallel in the terminal or remotely. The CLI provides functionality to read, modify, and delete files, and execute shell commands, with user approval required for these actions. This expands the reach of AI-powered code iteration beyond the Cursor editor.
Advertisement
Jan 2025A new Tab modelThe 'Fusion' model for Tab has been released, significantly improving cursor jump quality and edit suggestion length and accuracy. It achieves lower latency (260ms p50 vs 475ms) and supports longer context windows (13000 tokens vs 5500). Gains are attributed to cleaner data, longer context, training for larger edits, synthetic data, and inference performance engineering.
Jan 2025Character Prefix ConditioningIntroduced and described the algorithm for character prefix conditioning, which enables language models to generate code completions that precisely start with the user's typed characters, addressing the challenge of tokenization boundaries. This algorithm allows sampling from a distribution conditional on a character prefix, rather than a token prefix, by defining a new distribution q(s) = p(s | s starts with P) and outlining the process for autoregressive sampling from this distribution.
Sep 2024Iterating with shadow workspaces · CursorThis post details the design criteria for shadow workspaces, focusing on LSP-usability and runnability while maintaining independence, privacy, concurrency, universality, maintainability, and speed. It describes the current implementation using a hidden Electron window and the IPC mechanisms involved, including the use of gRPC and buf for communication between extension hosts. It also outlines the challenges with simpler solutions like copying TextModel objects and discusses future plans for a kernel-level folder proxy.
May 2024More problems · CursorThis post outlines future research directions for AI programming, including next action prediction, improving edit quality through scaled inference, optimizing context retrieval for codebases (including multi-hop context), and enhancing bug detection and debugging capabilities. It proposes research into action prediction models, advanced inference techniques for larger edits, novel retrieval methods for codebases, and improved debugging tools that track runtime information.
May 2024Editing Files at 1000 Tokens per SecondIntroduced a "fast-apply" model and "speculative edits" technique to achieve ~1000 tokens per second for code edits. This involved training custom models (e.g., Llama-3-70b-ft) using synthetic data and optimizing inference for full-file code rewrites, demonstrating significant speedups over vanilla inference and existing speculative decoding methods. The post details the evaluation methodology for prompted rewrites, the rationale for full-file rewrites over diffs, and the training pipeline for the specialized model. Future directions include long-context training and knowledge distillation.
Oct 2023Our problems · CursorThis post details several new problems and future directions for Cursor's AI-powered coding capabilities. It introduces the concept of a custom and fast reranker model to improve context retrieval for LLMs, aiming to filter down 500k tokens to the most relevant 8k. It also discusses the need for a 'copilot for edits' to handle small code changes, the development of constrained, in-flow agents for engineering tasks (starting with folder-level agents), and both passive and active bug-finding mechanisms. Furthermore, it addresses the challenge of enabling larger edits to files and directories, requiring smart model capabilities and real-time diff presentation. The post also highlights the significant scaling challenges in indexing, mentioning 1.4 billion vectors and 150 thousand codebases indexed, and the need for a custom indexing system. Future ideas include 'time warp', deep codebase understanding, 'reader mode', 'pseudo-code mode', and automated stack trace debugging.