Refactoring Legacy Codebases

The webinar focuses on refactoring legacy codebases using Cursor, an AI-powered development tool. Amrita, a field engineer at Cursor, leads the session with her colleague Reagan. The core objective is to demonstrate how AI, specifically Cursor, can streamline the process of modernizing old code, improve test coverage, and manage complex migrations. Cursor offers several ways to interact with its AI capabilities. The **Agents window** provides a side-panel interface where users can chat with AI agents, select models, and work across different repositories. For those who prefer a familiar environment, Cursor also functions as a fork of VS Code, retaining its debugging features, extensions, and terminal. Additionally, a **Cursor CLI** is available for users who prefer T-Max workflows or want to integrate Cursor with IDEs like Android Studio or Xcode. The most advanced and emphasized feature is **Cursor Cloud Agents**. These agents operate autonomously on remote machines, allowing users to close their laptops while the agents continue working on tasks like large-scale refactoring or migrations. The cloud agents can even provide proof of completion through videos, screenshots, and testing. Cursor Cloud agents can be accessed through various interfaces: the **Cursor browser**, Slack integration for automated checks and PR creation, and a **mobile app** (currently iOS only) for managing agents on the go. A key differentiator for Cursor is its **model flexibility**. It supports models from various frontier labs (like Anthropic) and open-source models, including its own in-house models like Composer and the latest Grok 4.6. The session highlights that Grok 4.6 is a powerful and cost-effective model, particularly for long-running tasks. Cursor's strength lies not just in offering these models but in enabling users to **combine different models** for optimal results at a lower cost. A case study of rebuilding SQLite from scratch demonstrated how an "agent swarm" using a combination of planning and execution-focused models was more cost-effective than using a single, high-end model. The "Cursor harness" is the underlying technology that makes these models effective. It comprises **tool execution, cache management, and dynamic context management**, among other features. This harness layer enhances the AI's capability for large-scale refactoring and migrations, such as upgrading SQL Alchemy versions. The webinar outlines a four-step process for refactoring legacy code with Cursor: 1. **Audit and Strategize with Cursor Canvas and Plan Mode:** * **Cursor Canvas** is a built-in tool that analyzes codebase data (files, spreadsheets, etc.) to create interactive visualizations. It helps audit code, identify test coverage gaps, and prioritize migration tasks. The canvas can be modified and shared, acting as a living document. * **Plan Mode** is used to develop a migration strategy. Users can define goals, and Cursor will generate a markdown document outlining the plan, technical architecture diagrams (using Mermaid, with integration options for Lucidchart, Excalidraw, and FigJam), sequence diagrams, API designs, and data model changes. Cursor actively prompts users for clarification during this phase to refine the plan. 2. **Leverage Plugins to Write Tickets:** * Cursor integrates with various **Managed Cloud Platforms (MCPs)** like Atlassian (Jira), Linear, monday.com, Google Drive, Google Calendar, and Gmail. * For the demo, Cursor is used to convert the migration plan into Jira tickets, creating an epic for a PHP to React migration. This leverages the Atlassian MCP. 3. **Hand Off to Cursor Cloud Agents:** * Once the plan is broken down into tickets, these can be delegated to **Cursor Cloud Agents**. These agents operate on remote VMs, eliminating the need for local machine resources. * Users can connect multiple repositories to a cloud agent, which is beneficial for microservices architectures or when changes in one service impact others. * Cloud agents offer the same model flexibility as local agents, but users are encouraged to use cheaper, faster models for cost efficiency on long-running tasks. * A key feature of cloud agents is their ability to perform **self-verification** using screenshots and videos of their work, ensuring no regressions are introduced. They can also interact with the operating system and perform end-to-end testing on full-stack applications. * The process involves setting up a remote environment (which takes longer the first time but is cached for subsequent runs), and then the agent executes the assigned tickets, creating separate Pull Requests (PRs) for each ticket to facilitate easier review. 4. **Automations for Maintenance and Stability:** * **Automations** are scheduled or triggered cloud agents designed for ongoing maintenance and proactive problem-solving. * They can be triggered by time, events (like PR creation), or messages (Slack/Teams). * Examples include a "feature flag cleaner" that weekly scans for stale feature flags, cross-references with tools like DataDog and Sentry, and automatically opens a PR to remove them. * Automations can also be used for dependency updates, security scans, and incident triaging, helping to prevent the accumulation of technical debt and the need for future large-scale refactoring. * Cursor provides templates for common automation tasks, making it easier to get started. The demo showcases migrating components from the WordPress codebase from PHP to React. The process involved using `/canvas` to audit test coverage and `/plan` mode to strategize the migration. The plan was then converted into Jira tickets using the Atlassian plugin. These tickets were then handed off to a cloud agent, which was configured to work across multiple repositories if needed. The agent was set to create individual PRs for each ticket and verify its work with screenshots and videos. The session also touched upon the **mobile app**, which allows users to monitor and manage their agents remotely. Furthermore, the concept of **skills** was discussed, highlighting that plugins often come with pre-built skills, and users can create their own custom skills for personalized workflows. For security and private network access, Cursor Cloud agents support various methods like Cloudflare Tunnels, AWS Private Link, and VPNs. They can also interface with infrastructure via private connectivity options detailed in Cursor's documentation. Regarding learning and adaptation, Cursor is developing **memory management** features. The `continual learning` plugin allows agents to learn from user interactions and coding styles, improving future performance. In automations, a `memories.mmd` file helps agents learn from past runs and feedback, becoming more effective over time. The webinar concluded by emphasizing the power of Cursor Cloud Agents for asynchronous work, distributed teams, and proactive code maintenance. It encouraged users to explore Cursor Cloud via `cursor.com/onboard` and leverage the provided documentation and cookbook for further guidance. The session also briefly touched upon Grokbot, a general knowledge product that can be used for coding and integrates local and cloud environments. The discussion addressed how cloud environments handle environment variables and secrets, which are managed during the cloud agent setup. It also covered how cloud agents can access local resources through specific network configurations like private links or tunnels, and that plans or skills committed to a repository can be accessed by automations. Finally, it was clarified that while local conversations can be referenced, the ability to reference past cloud conversations directly within the cloud environment would require further confirmation.