February 2025 Summaries

Cursor's Agent mode, introduced in November 2024, revolutionizes coding by enabling autonomous capabilities such as reading terminal outputs, fixing bugs, and refactoring code using the E2B Sandboxes and AgentKit. This implementation focuses on building a Coding Agent that replicates Cursor's functionalities, including generating projects, running tests, and creating or updating files in a sandbox environment. The Coding Agent relies on the ReAct and CoT prompt structures and uses tools like createOrUpdateFiles and readFiles to interact with the E2B Sandbox filesystem. By leveraging the AgentKit library, the agent operates within an "autonomous loop," allowing it to refine its actions based on tool outputs until a task is complete. This design, which emphasizes "LLM empathy," mirrors the capabilities of Cursor's Agent, and by integrating Model Context Protocol (MCP) servers, the agent can also browse the web autonomously. Consequently, the Coding Agent can perform complex tasks such as creating a Next.js TodoList with unit tests, showcasing its potential to simplify and enhance coding workflows.

OverlayFS is a Linux filesystem that efficiently manages multiple Firecracker instances by layering one filesystem over another, creating a merged view without duplicating data. This technique uses a read-only lower directory containing the root filesystem and a writable upper directory for modifications, employing copy-on-write (COW) to maintain the integrity of the base system. By leveraging OverlayFS, users can significantly reduce storage requirements and avoid copying hundreds of megabytes for each instance, which is especially beneficial when running thousands of instances. The blog post explains how to set up OverlayFS with Firecracker, detailing the creation and configuration of Squashfs images for read-only filesystems and ext4 images for writable overlays. Additionally, it demonstrates configuring Firecracker to utilize OverlayFS, allowing for scalable and efficient instance management without the need for persistent storage, making it ideal for scenarios requiring numerous small instances.

UC Berkeley’s Arena research team created WebDev Arena, a free open-source platform where large language models compete to build web applications and users vote on their results, producing a public leaderboard. Evaluating such models requires secure, fast, simultaneous execution of potentially complex generated code, while preventing interference or system access between competing models. To address these requirements, Arena integrated E2B in November 2024, using isolated virtual-machine sandboxes that start in roughly 150 milliseconds and support diverse frameworks, dependencies, and models including Claude, GPT-4, Gemini, DeepSeek, and Qwen. The implementation supports configurable sandbox creation, dynamic package installation, real-time dependency logs, and separate modes for code interpretation and web-app deployment. Arena reported that more than 50,000 users launched over 230,000 sandboxes, while WebDev Arena has completed over 50,000 model comparisons, with Claude 3.5 Sonnet leading the leaderboard ahead of DeepSeek-R1; future plans include adding models and expanding into broader software-engineering and general-purpose coding evaluations.