March 2026 Summaries

Composer 2, a coding model for agentic software engineering, is detailed in a technical report that outlines its training process and infrastructure. The model undergoes two phases of training: continued pretraining on a code-focused data mix to enhance the base model's coding knowledge and large-scale reinforcement learning (RL) aimed at improving agent performance in realistic Cursor sessions. A notable improvement in both average and best-of-K performance indicates that Composer 2 learns new solution paths rather than relying solely on known ones. Real-world evaluation is conducted using CursorBench, a benchmark built from genuine coding tasks, which ensures alignment with practical problems faced by developers. Composer 2 achieves a 61.3 score on CursorBench, a 37% improvement over its predecessor, and performs competitively on public benchmarks while maintaining lower inference costs. The training required extensive infrastructure development, including custom low-precision kernels, asynchronous RL pipelines, and the Anyrun compute platform, with a focus on efficient and scalable model training. The report also includes details on weight synchronization, fault tolerance, and environment fidelity, acknowledging the contributions of collaboration partners and the broader open-source community.

Real-time reinforcement learning (RL) is being leveraged to enhance coding models like Composer by using actual user interactions as training signals, thereby addressing the train-test mismatch often encountered in simulated environments. This approach involves frequent deployment of improved model versions, thanks to an infrastructure that translates user feedback into reward signals, allowing updates as often as every five hours. Despite the potential for reward hacking, where models exploit flaws in the reward system, real-time RL incorporates user feedback to refine the training process and mitigate such risks. The system's design enables continuous improvement by learning from longer, more complex user interactions and allows for specialization based on specific organizational needs. This method ensures that the model's training data remains on-policy, reducing the likelihood of over-optimization and enhancing model performance in real-world applications.

Cursor has introduced self-hosted cloud agents, providing enterprises with the ability to maintain tighter security while leveraging the capabilities of cloud agents. These agents operate in isolated virtual machines within an organization's own network, ensuring that code, tool execution, and build artifacts remain secure and compliant with internal requirements. Self-hosted agents are particularly beneficial for companies with complex development environments, as they can access necessary caches, dependencies, and network endpoints. This solution allows engineering teams to focus more on development rather than infrastructure maintenance, by delegating software build tasks to Cursor's cloud agents. The agents can autonomously clone repositories, set up environments, write and test code, and push changes, all while remaining compliant with the organization's security protocols. By using a combination of a worker process and Cursor's agent harness, these agents can execute tasks efficiently and scale as needed, which is particularly useful for large teams and organizations operating in regulated industries.

The evolution of search tools for programming has seen significant advancements since the introduction of grep in 1973. Initially a basic utility for matching regular expressions, grep has been surpassed by more sophisticated indexing tools and integrated development environments (IDEs) that use syntactic and semantic indexes to navigate codebases efficiently. The emergence of the Language Server Protocol (LSP) standardized this functionality across different editors. However, with the advent of Agentic coding, there is renewed interest in grep-style searches, particularly using ripgrep, a faster alternative due to its efficient handling of regular expressions. Despite its speed, ripgrep struggles with large codebases, prompting innovations in indexing methods such as inverted indexes, trigram decomposition, suffix arrays, and sparse n-grams, each offering various trade-offs between speed, storage efficiency, and query specificity. The latest efforts focus on implementing these indexes client-side to minimize latency and improve performance in large enterprise repositories, emphasizing the importance of efficient regular expression searches in modern development workflows.

Composer 2, now available in Cursor, is a cutting-edge coding model that offers significant improvements over its predecessors, achieving higher scores on benchmarks such as Terminal-Bench 2.0 and SWE-bench Multilingual. Priced at an economical $0.50 per million input tokens and $2.50 per million output tokens, it provides an optimal balance of intelligence and cost. The advancements in Composer 2 stem from its first continued pretraining run, which enhances its base for scaling reinforcement learning, allowing it to handle complex tasks that require numerous actions. Additionally, a faster variant is available at a higher price, making rapid processing the default option. The model's performance was validated using the Harbor evaluation framework, and its scores were compared with those of other models using standardized benchmarks. Composer 2 can be explored further in Cursor or the early alpha of a new interface, with detailed documentation accessible for users.

Money Forward, a financial services company in the Asia-Pacific region, has significantly improved its software development lifecycle by implementing Cursor, an AI-driven tool that enhances productivity across various teams, including engineering, product, design, and quality assurance (QA). Initially, the engineering team experienced substantial time savings of 15–20 hours per week per engineer by using Cursor for tasks such as refactoring service layers, optimizing applications, and managing deployments. This success led to a company-wide adoption, with over 1,000 employees now utilizing Cursor daily. The tool's model-agnostic infrastructure and visual capabilities have facilitated seamless integration across departments, allowing designers to prototype directly against live frontends and QA engineers to automate test generation, reducing the time spent on these tasks by 70%. Product managers have also benefited from Cursor's ability to extract system relationships and generate architecture diagrams, leading to more efficient product requirement development. This widespread adoption has resulted in enhanced collaboration and efficiency, allowing teams to focus on high-value tasks and improve overall software quality.

Composer, a specialized model designed for long-horizon tasks, employs a reinforcement learning process called self-summarization to improve its performance on complex coding challenges. This approach allows Composer to handle tasks that require extensive sequences of actions by summarizing its context when reaching a fixed token-length trigger, thus overcoming the limitations of compaction techniques that can cause loss of critical information. By integrating self-summarization into its training, Composer can efficiently condense context into high-value summaries with fewer tokens, significantly enhancing its performance in context-constrained environments. Testing against a baseline, Composer demonstrated superior results, reducing compaction errors by 50% while requiring only a fraction of the tokens. This capability enables Composer to tackle intricate problems, such as those in the Terminal-Bench 2.0, by condensing over 100,000 tokens into concise, actionable information. The ongoing development of Composer aims to extend its applicability to even more complex tasks, including multi-agent coordination, promising advancements in the field of agentic systems.

Over the past nine months, the implementation of Cursor Automations has significantly increased the efficiency of security processes, allowing for the creation of security agents that autonomously identify and resolve vulnerabilities in the codebase. Four new automation templates have been released to enable other security teams to customize and build agents that automatically address a variety of security concerns. The automation architecture includes features such as integrations for webhooks, GitHub pull requests, and codebase monitoring, as well as a robust agent environment powered by cloud agents. A key component is the security MCP tool, deployed as a serverless Lambda function, which manages data persistence, deduplication of findings, and consistent reporting through Slack. Among the implemented automations are Agentic Security Review, which focuses on PR security findings, Vuln Hunter, which scans existing code for vulnerabilities, Anybump, which automates dependency patching, and Invariant Sentinel, which monitors code for security and compliance drift. These automations have streamlined security processes, leading to improved detection and prevention of issues, with plans to extend their functionality to areas like vulnerability reporting and privacy compliance monitoring.

Cursor has expanded its Marketplace by introducing over 30 new plugins from notable partners like Atlassian, Datadog, GitLab, and more, enhancing its ability to integrate with various tools and improve workflow efficiency. These plugins enable Cursor to perform actions across different platforms, allowing users to install or develop custom plugins for broader functionality. Particularly beneficial for Teams and Enterprise plans, admins can create private team marketplaces for managing these plugins. The new plugins enhance the development workflow by empowering agents to operate independently with improved infrastructure and productivity tools, such as querying logs with the Datadog plugin and managing project tasks with the monday.com plugin. These tools also support automation through Cursor Automations, allowing agents to run tasks based on schedules or events using custom instructions.

Cursor has developed CursorBench, an internal evaluation suite, to assess the performance of coding agents more accurately than public benchmarks. Built on real sessions from their engineering team, CursorBench measures various dimensions of agent performance such as solution correctness, code quality, efficiency, and interaction behavior, making it more aligned with real-world developer outcomes than traditional benchmarks. The suite addresses the limitations of public benchmarks, which often fail to differentiate between models effectively due to issues like misalignment with actual coding tasks, narrow grading criteria, and contamination from training data. CursorBench's tasks are derived from actual developer queries and solutions, ensuring they are relevant and challenging, with the scope of correctness evaluations doubling since its inception. By combining online and offline evaluations, Cursor ensures that model quality aligns with developers' practical experiences, allowing for the identification of regressions that offline methods might miss. As development work evolves to involve long-running agents, Cursor plans to adapt CursorBench to maintain its relevance and effectiveness, aiming to continue improving the agent experience in production environments.

Cursor Automations introduces a powerful tool for building always-on agents that streamline various aspects of the software development lifecycle by executing tasks on schedules or in response to specific events such as Slack messages, GitHub PRs, or PagerDuty incidents. These automations enhance productivity by enabling flexible configurations, including custom webhooks, and are used for code review, security audits, incident response, and routine chores like summarizing weekly changes or ensuring test coverage. By leveraging cloud agents, teams can automate repetitive tasks, allowing engineers to focus on more critical work, as exemplified by Rippling's use of automations for tasks like incident triage and status reporting. The ability to configure agents for continuous monitoring and improvement of the codebase allows companies like Cursor and Runlayer to operate more efficiently, akin to having a software factory that continuously optimizes production processes.

Cursor is now integrated into IntelliJ IDEA, PyCharm, WebStorm, and other JetBrains IDEs through the Agent Client Protocol (ACP), allowing developers to use any frontier model for agent-driven development. This integration enables developers to choose from models provided by OpenAI, Anthropic, Google, and Cursor, with a custom-built agent harness to enhance output quality and performance. Cursor uses secure codebase indexing and semantic search to manage large enterprise codebases, and its integration with JetBrains IDEs combines these capabilities with deep code intelligence. This collaboration is designed to benefit developers by providing powerful AI assistance while maintaining control over their development environment, and it marks the beginning of further collaborations between Cursor and JetBrains to enhance agentic coding capabilities.

Cursor's integration into customer support workflows has significantly improved efficiency by consolidating code, logs, team knowledge, and past conversations into a single session, enhancing support engineer throughput by 5–10 times. The process begins with investigating from the codebase using Ask Mode, which allows tracing through product behavior and indexing across multiple repositories within a multi-root workspace. By utilizing MCP servers, support engineers can access comprehensive context from databases, event logs, communication platforms, engineering tickets, and internal documentation, all within Cursor. When issues arise, Datadog MCP provides relevant logs for error analysis, and the integration with Slack and other platforms helps track similar cases. Determining whether an issue is a bug involves cross-referencing with runbooks pulled through Notion MCP, while Linear MCP aids in filing detailed bug reports. Documentation updates are streamlined by tagging Cursor in Slack, prompting cloud agents to update the docs repository. Automation through slash commands, rules, skills, and subagents allows parallel execution of common steps, further improving productivity. This AI-native approach to technical support has enabled a small team to effectively manage a rapidly growing user base by reducing the need to switch between tools and teams.

PlanetScale, a company managing cloud database workloads, has integrated Bugbot into its software development lifecycle to enhance code review and maintain product reliability. As code generation became rapid and inexpensive, the bottleneck shifted to code review, prompting the need for a dedicated agentic review layer. Bugbot effectively identifies complex issues like state synchronization gaps and logical flow errors that human reviewers often miss, ensuring these are addressed before code merges. This adoption has led to a reported 80% resolution rate of Bugbot's findings, saving the equivalent of two full-time engineers' worth of review effort and allowing PlanetScale to ship code with greater confidence. Bugbot's ability to surface critical issues has made it indispensable, as it allows engineers to focus on complex infrastructure problems rather than manual code reviews, thereby accelerating software delivery without compromising quality.