February 2026 Summaries

Cursor is transitioning from traditional keystroke coding to a new era of AI-assisted development, where autonomous agents handle larger and more complex tasks with minimal human intervention. Initially, tools like Tab facilitated AI-assisted coding by automating repetitive tasks, but the introduction of advanced agents has shifted developer roles from writing code to managing software creation processes. These agents operate in the cloud, allowing for parallel task execution and enabling developers to focus on problem definition and output review. This shift has led to a significant increase in agent usage within Cursor, with over one-third of internal pull requests now generated by these cloud-based agents. As this approach gains traction, challenges such as ensuring reliable test environments and providing agents with the necessary tools and context remain, but the launch of Cursor's cloud agents marks a pivotal step towards standardizing this method in software development.

Bugbot, a code review agent, has been enhanced to autonomously identify and fix issues in pull requests (PRs) through its new Autofix feature, which operates via cloud agents in virtual machines. This feature, now available to all users after a successful beta phase, has improved the efficiency of Bugbot by increasing the number of identified issues and the resolution rate, now at 76%. Over 35% of Autofix changes are successfully merged into base PRs, demonstrating its effectiveness. Looking ahead, Bugbot aims to expand its functionality to include customizable automations for broader workflows beyond code review and to enhance its capabilities in verifying findings, conducting in-depth research on complex issues, and continuously scanning codebases for bug resolution. Users can enable Bugbot Autofix through their dashboard or consult further documentation to learn more.

Cloud agents are transforming software development by providing self-contained virtual environments where they autonomously create and test code, thus enhancing productivity and minimizing resource conflicts. Cursor's new version of cloud agents, accessible through various platforms like web, mobile, and Slack, allows these agents to integrate into codebases, produce merge-ready pull requests, and interact with software directly in isolated sandboxes. This shift enables agents to handle more ambitious tasks independently, such as building new features, reproducing vulnerabilities, and performing UI testing, while developers focus on strategic decision-making. By leveraging cloud agents, Cursor has achieved a significant change in software development practices, with agents now responsible for over 30% of the merged pull requests. The future aims to achieve self-driving codebases where agents manage end-to-end processes, requiring advancements in tooling, models, and interaction patterns to coordinate agent activities effectively.

Coding agents have become increasingly adept at executing terminal commands, raising concerns about security risks when users auto-approve their actions, potentially leading to errors such as data loss or code issues. To address these challenges, a sandboxing solution has been implemented across macOS, Linux, and Windows, allowing agents to operate within a controlled environment and only seek approval when necessary, particularly for internet access, which significantly reduces interruptions and approval fatigue. The sandboxing approach varies by platform, utilizing tools like Seatbelt on macOS, Landlock and seccomp on Linux, and WSL2 on Windows, with efforts underway to create better native support for Windows. Enhancements to the agent harness ensure agents can navigate sandbox constraints effectively, improving their ability to manage permissions and operate efficiently. The successful rollout of sandboxing has seen widespread adoption, including by major enterprise clients, highlighting the importance of execution boundaries for agents as they transition from code generation to production system operations.

Cursor has introduced a plugin system that enhances its capabilities by allowing agents to integrate with external tools and acquire new functionalities, covering the entire product development lifecycle. These plugins, sourced from partners like Amplitude, AWS, and Stripe, enable various functions such as deploying services, implementing payment systems, and conducting advanced testing. Users can explore and install prebuilt plugins from the Cursor Marketplace or create and share their own, utilizing components like skills, subagents, and MCP servers to execute tasks efficiently. The platform aims to streamline development processes by integrating with existing tools, allowing seamless orchestration of tasks, from planning and design to data analytics and infrastructure management. Furthermore, Cursor is developing private team marketplaces to facilitate the secure exchange of plugins within organizations.

Stripe, a leader in developer experience with a reputation for 99.9999% API uptime, has swiftly integrated AI coding tools like Cursor into its developer infrastructure, impacting its 3,000-plus engineers. Scott MacVicar, Head of Developer Infrastructure at Stripe, emphasizes the importance of rapid AI adoption, noting that they equipped engineers with preconfigured tools, including Cursor, from day one to maximize productivity. To maintain quality amid increased coding velocity, Stripe adapted its code review process, utilizing large language models (LLMs) to highlight complex areas needing attention. Interestingly, the most experienced engineers have experienced significant productivity boosts using AI, contrary to expectations that junior engineers would benefit most. MacVicar suggests that new hires might benefit from familiarizing themselves with the codebase before using AI tools, highlighting the evolving nature of AI integration and the need for engineering leaders to begin experimenting with these technologies to optimize their use.

Box, a leading company in content and data management, has significantly enhanced its product development process by adopting Cursor, an enterprise-ready agentic coding platform. This strategic move has led to a 30-50% increase in product roadmap throughput and an 80-90% faster migration to React, thanks to Cursor's robust data privacy and security controls. Over 85% of Box's engineering organization now uses Cursor, facilitating a seamless integration that allows developers to set agent guardrails directly in code and improve automated test coverage. This adoption has not only accelerated project completion but also reinforced Box's commitment to building secure, high-quality enterprise-grade products. The introduction of a mentorship program further boosted Cursor's usage, fostering a culture of innovation and efficiency within the organization. With Cursor embedded across the software development lifecycle, Box has empowered its developers to focus on delivering greater value to customers while maintaining stringent security and quality standards.

Cursor has unveiled a research preview of its long-running agents, designed to autonomously tackle complex projects, available to Ultra, Teams, and Enterprise users. These agents, developed from research on autonomous web browser building, overcome limitations of frontier models on long-horizon tasks by employing a custom harness that emphasizes planning and follow-through, resulting in larger, production-ready pull requests (PRs) with minimal follow-up. The agents successfully completed tasks like building new platforms, refactoring systems, and optimizing performance, demonstrating capabilities beyond those of synchronous agents by offering a more comprehensive approach to problem-solving. Early participants found the agents reduced timelines from several months to days, allowing parallel project handling. The initiative marks a step toward autonomous codebases, enabling agents to manage larger tasks with less human intervention, and prompting the development of new tools to handle increased code generation and deployment safely.

Cursor has introduced two usage pools and increased limits for Auto and Composer 1.5 to better support developers' evolving needs, particularly those coding with agents. Composer 1.5, a model trained by Cursor, offers three times the usage limit of its predecessor and is temporarily increased to six times until February 16, providing a balance of speed, intelligence, and affordability. The update aims to accommodate developers who either opt for the latest models or seek a cost-effective alternative, and includes a new page in the editor for enhanced usage visibility. The Composer 1.5 model, evaluated using Terminal-Bench 2.0, demonstrates significant capability, allowing Cursor to provide more usage sustainably. This update applies to all individual plans, with usage limits resetting monthly, and offers a way to monitor limits through the editor, ensuring users can make informed decisions about their usage.

Composer 1.5, the latest version of the agentic coding model, represents a significant advancement in coding ability by enhancing both speed and intelligence for everyday use. Developed by scaling reinforcement learning 20 times further on the same pretrained model, Composer 1.5 surpasses its predecessor in performance, particularly on challenging coding tasks, as evidenced by internal benchmarks. It introduces "thinking tokens" that allow the model to reason through a user’s codebase and plan the next steps, balancing quick responses for simpler problems with more thoughtful consideration for complex ones. Composer 1.5 also features self-summarization capabilities, enabling it to maintain accuracy as context length varies by generating useful summaries when the available context runs out. This recursive self-summarization process contributes to its robustness, making Composer 1.5 a highly recommended tool for interactive coding tasks, showcasing the potential of scaling reinforcement learning for continuous intelligence improvements in coding models.

NVIDIA has integrated Cursor into its software development lifecycle (SDLC) to enhance AI utilization and streamline various processes, resulting in improved engineering velocity and onboarding efficiency for new developers. With over 30,000 active users, Cursor has tripled the amount of committed code, highlighting its effectiveness in code generation, testing, reviews, and debugging. NVIDIA has extended Cursor's capabilities beyond individual productivity to automate core workflows, such as git flow and bug fixes, through custom rules and semantic reasoning over large codebases. This has maintained code quality while accelerating development speed, enabling developers to bridge skill gaps and contribute more quickly to unfamiliar codebases. The adoption of Cursor has made software development more engaging for engineers and significantly transformed NVIDIA's engineering practices.

Research on scaling long-running autonomous coding has led to the development of a new agent harness that orchestrates thousands of agents in parallel to build a web browser, marking a significant milestone in autonomous software development. Initially, attempts to use a single agent for complex tasks like building a browser were unsuccessful due to task overwhelm, prompting the transition to a multi-agent system with structured roles for planning, executing, and evaluating tasks. This structure improved coordination but revealed challenges with synchronization and resource management. The final system design incorporates recursive planning, where a root planner delegates tasks to subplanners and workers, enhancing throughput and maintaining dynamism without global synchronization overhead. The research highlighted the importance of precise initial instructions, constraints over instructions, and empirical data-driven adjustments to optimize agent performance. The system achieved notable throughput, peaking at around 1,000 commits per hour, while accepting a manageable error rate to maintain efficiency. This research not only advances the understanding of autonomous coding systems but also reflects emergent behaviors akin to current software team structures, suggesting potential future directions for long-running AI agents.