April 2026 Summaries

In the development of the Cursor agent harness, a vision-driven process is used to create a robust software product by forming hypotheses, running experiments, and iterating through feedback from evaluations and real usage. The harness is optimized to match model strengths and includes dynamic context management, moving away from static context and strict guardrails as models improve. Online and offline tests, including public benchmarks and A/B testing, assess changes to the harness, focusing on metrics like latency and agent-generated code quality. The system is designed to detect and repair errors, with specific classifications for expected and unknown errors, and utilizes automated tools to handle issues efficiently. The harness is customized for different models, ensuring compatibility with model-specific tool formats and prompting methods. Challenges such as mid-chat model switching are addressed through custom instructions and conversation summarization. The harness is envisioned as key to the future of AI-assisted software engineering, where multiple specialized agents collaborate within a coherent workflow.

The Cursor SDK has been introduced to enable developers to build coding agents using the same runtime, harness, and models that power Cursor's desktop, CLI, and web applications. This SDK allows for the deployment of agents without the complexities of building an entire agent stack, offering the flexibility to run agents locally, on self-hosted servers, or in the cloud. With features like secure sandboxing, session management, and context management, the SDK simplifies the process of building efficient and reliable agents. It supports various models, including the cost-effective Composer 2, allowing users to choose the most suitable model for their tasks. The SDK is currently in public beta and is being used by teams to automate workflows directly from CI/CD pipelines, integrate agents into core products, and enhance customer-facing applications. Additionally, the Cursor SDK is designed to run on a production-ready cloud infrastructure, making it easier to manage agents without the overhead of VM management, and it includes starter projects and tools to help teams quickly deploy and iterate their solutions.

National Australia Bank (NAB) has implemented the Cursor platform across its development teams to enhance project delivery speed and quality, after evaluating and selecting it over Amazon Q and GitHub Copilot. This strategic move has enabled NAB to expedite legacy codebase modernizations, such as transitioning from monolithic structures to microservices and migrating from Assembly-based mainframes, achieving a threefold increase in speed. The flexibility of Cursor's model allows engineers to choose different models based on task complexity, while its understanding of the codebase across various tech stacks ensures accurate agent behavior. NAB has also benefited from Cursor's extensibility, using it to enforce development standards via a custom internal library. The bank plans to embed Cursor into all phases of the software lifecycle, including code review and quality assurance, while also democratizing software development across different organizational roles. Notable achievements include completing a hardware-agnostic payment app in three weeks, significantly faster than the projected four-month timeline, and accelerating the refactoring of a business lending application from Silverlight to Java microservices, with all pre-development tasks completed in just one week.

Cursor, a desktop app built on Visual Studio Code and Electron, has faced stability challenges as its user base and features have expanded, with crashes primarily caused by out-of-memory (OOM) issues. To address these, the team has implemented systems for monitoring crashes and memory pressure, resulting in a significant reduction in OOM incidents. The app's architecture, featuring multiple processes for editors and utility functions, makes it susceptible to crashes, which are reported via telemetry. Cursor employs dual debugging strategies: a top-down approach focusing on memory-intensive features and a bottom-up approach tracing individual crash events to their root causes. By using crash stacks, heap snapshots, and continuous heap allocation profiling, the team can identify and mitigate acute and slow-and-steady OOM patterns. They have also implemented process isolation for extensions to limit the impact of crashes and are committed to preventing regressions by enhancing developer awareness and utilizing automated testing tools. The company continues to prioritize application stability while embracing the challenges of agentic software development.

Cursor's partnership with SpaceX aims to significantly enhance their model training efforts by utilizing xAI's Colossus infrastructure, addressing previous limitations imposed by compute constraints. Since the release of Composer less than six months ago, Cursor has achieved substantial advancements in their agentic coding model, with Composer 1.5 increasing reinforcement learning capabilities by over 20 times and Composer 2 introducing continued pretraining that delivers frontier-level performance at a reduced cost. This collaboration is expected to further scale the intelligence of Cursor's models, building on the successes of previous iterations.

Cursor has introduced a new feature allowing agents to create visual canvases that facilitate the representation of data and information interactively, enhancing the usability of the platform by moving beyond traditional text-based interfaces. These canvases utilize a React-based UI library with components such as tables, diagrams, and charts, enabling agents to tailor custom interfaces for tasks like reviewing pull requests, managing data-intensive operations, or conducting evaluations. The introduction of canvases has proven particularly effective in areas like incident response and eval analysis, where they allow for the integration of data from various sources into comprehensive visualizations, aiding in the identification of patterns and insights that might otherwise be overlooked. This innovation is part of a broader effort to improve human-agent collaboration through increased information bandwidth, supported by additional features like Design Mode and enhanced voice input, with the latest updates available in Cursor 3.1.

The study conducted in collaboration with Professor Suproteem Sarkar from the University of Chicago examines how advancements in AI models, specifically the releases of Opus 4.5 and GPT-5.2, impact developer work habits across 500 companies from July 2025 to March 2026. The research highlights a significant increase in AI usage, marked by a 44% rise in average weekly messages per user, which aligns with a Jevons-like effect where efficiency gains lead to increased consumption. Initially, developers used the improved models to perform more tasks of similar complexity, but after 4–6 weeks, they began tackling more complex tasks, with high-complexity messages growing by 68%. The shift in task distribution indicates a move towards managing AI-generated output, with substantial increases in documentation, architecture, code review, and learning, while simpler tasks saw less growth. The study suggests that while AI facilitates existing work, it also opens up new productive opportunities, particularly in industries like finance, media, and advertising, where competitive pressures and new opportunities drive adoption.

Amplitude's engineering team has significantly enhanced its software development process by integrating Cursor's cloud-based automation agents, which streamline the journey from feature ideas to production with minimal developer intervention. These agents autonomously manage a range of tasks, from investigating customer-reported issues in Slack to opening tickets and implementing code fixes, thereby speeding up response times and reducing manual workload. The transition from local to cloud agents has allowed for better parallelism, autonomy, and resource management, resulting in a threefold increase in weekly production commits. The implementation of Bugbot, an agentic review layer, has improved code review efficiency by automatically handling low-risk pull requests and accurately identifying complex bugs. Additionally, Amplitude has automated legacy code refactoring using cursor automations, which continuously run in the background, thus optimizing the codebase without interrupting other tasks. With these advancements, Amplitude aims to further extend automation into the latter stages of the development lifecycle, including CI/CD pipelines and deployment, to achieve a fully autonomous development pipeline.

Researchers have developed a multi-agent system capable of autonomously building, maintaining, and deploying complex software, which they tested by optimizing CUDA kernels crucial for AI model training and inference on NVIDIA GPUs. In collaboration with NVIDIA, the system tackled 235 optimization problems and achieved a 38% geometric mean speedup by building and optimizing Blackwell GPU kernels from scratch, demonstrating the potential to significantly enhance GPU performance, reduce energy consumption, and lower costs. This accomplishment, typically requiring months or years from experienced kernel engineers, was achieved in weeks, indicating the system's ability to address complex, open-ended optimization problems by exploring a broader solution space beyond traditional manual methods. Using SOL-ExecBench for problem generation and benchmarking, the system exceeded baseline performance on 63% of problems and delivered over 2x improvements on 19% of them. The experiment highlighted the system's adaptability in employing distinct optimization strategies across various real-world constraints, suggesting that multi-agent architectures could soon become the standard in software development to address novel challenges that exceed current training data distributions.

Bugbot, an AI code review tool launched out of beta in July 2025, has demonstrated a significant improvement in bug resolution rates, achieving a resolution rate nearing 80% compared to other AI code review products. Initially relying on offline experiments for enhancements, Bugbot now leverages real-time feedback from its reviews to create learned rules that improve its performance. These learned rules are developed by analyzing developer reactions, replies to Bugbot’s comments, and human reviewer comments, with the ability to promote or disable rules based on their effectiveness. Since the introduction of learned rules, over 110,000 repositories have enabled this feature, resulting in more than 44,000 rules being generated. Bugbot's new capability allows it to continuously evolve by integrating feedback from natural experiments during code review processes, aiming to catch real bugs with a deeper understanding of unique codebases and team priorities. Users can manage Bugbot's learning process through the Cursor Dashboard, enhancing the tool's adaptability and effectiveness in code review tasks.

Warp decode is a novel approach to token generation in mixture-of-experts (MoE) inference systems, particularly on Blackwell GPUs, that enhances both performance and accuracy by reorienting the parallelism around outputs instead of experts. This method resulted from exploring the maximum achievable memory bandwidth for MoE decode on Blackwell, leading to the assignment of each warp to a single output value, which significantly improves throughput and accuracy, achieving a 1.84x throughput improvement and outputs 1.4x closer to the full FP32 reference. The approach eliminates the overhead of traditional expert-centric methods by removing unnecessary stages and intermediate buffers, allowing for more efficient scheduling and latency hiding. Warp decode is particularly advantageous in scenarios where there isn’t enough shared work per expert to justify traditional overheads, thus accelerating the research and training pipeline for Composer, a machine learning system, by allowing for faster model iteration and deployment. This strategy maintains hardware efficiency by sustaining substantial memory throughput and ensuring correctness across batch sizes, although it is not a universal replacement for expert-centric execution in scenarios like prefill and large-batch inference.

Cursor is evolving to meet the demands of the third era of software development, where agents autonomously handle much of the coding process. The introduction of Cursor 3 offers a unified workspace designed to facilitate collaboration between human developers and multiple agents across different repositories. This new interface enhances productivity by allowing seamless transitions between local and cloud environments and provides tools for managing agent sessions and code changes efficiently. With a focus on agent-first capabilities, Cursor 3 integrates new features such as a simplified user interface for handling pull requests, a built-in browser for local navigation, and a marketplace for plugins that extend agents' functionalities. By building from scratch rather than extending existing platforms, Cursor aims to redefine software development with AI-driven approaches, ultimately striving toward more autonomous and collaborative coding environments.