February 2024 Summaries

Pro users can now customize the number of vCPUs and RAM for their sandboxes through a newly dedicated page in the documentation, allowing any combination of CPU and RAM as outlined in provided tables. To achieve this customization, users must build a custom sandbox template using the E2B CLI by running the command `e2b template build` and specifying the `--cpu-count` and `--memory-mb` options. Those interested in using E2B sandboxes for running AI applications or agents are encouraged to start with this setup, and support is available via email at [email protected].

Support for Huge Pages has been added to sandboxes, significantly accelerating the start of memory-intensive tasks by up to five times. This improvement reduces the time needed to read 0.5 GB of data from 4.9 seconds to just 0.76 seconds. Huge Pages allow for the use of 2 MiB memory chunks instead of the default 4 KiB, minimizing the overhead from frequent memory requests by allowing for fewer, larger requests. This feature is enabled by default in newly created custom sandboxes, while existing ones require a rebuild using the e2b build command in the directory with the e2b.toml file. Users interested in trying out the sandboxes can begin immediately and seek support via [email protected] if needed.

An open-source cookbook example of a code interpreter using AutoGen agents was recently developed by community contributor Keegan McCallum, founder of Xler.ai, which is a multi-agent platform offering services like evaluation and deployment. This example project executes LLM-generated code in the cloud using E2B sandbox, a secure, long-running cloud environment that mirrors local execution and supports various LLMs, including GPTs and Claude. E2B, fully open-sourced, provides an infrastructure layer for running AI applications securely, offering a solution to the limitations and risks of local execution via Docker. Users can explore E2B sandboxes for free through its documentation, and the E2B cookbook encourages contributions of LLM-powered code interpreters or coding AI agents that utilize E2B sandboxes.

The paper "More Agents Is All You Need" suggests that the performance of Large Language Models (LLMs) improves with the number of agents, a concept that supports the growing popularity of multi-agent frameworks like CrewAI and AutoGen. CrewAI, based on LangChain, orchestrates multiple agents working on user-defined tasks and allows delegation among them, making it quick to set up for various applications such as stock analysis or generating Instagram posts. AutoGen, on the other hand, excels in executing LLM-generated code, typically using Docker containers, which might limit some use cases but offers a cloud alternative for safer execution. Both frameworks have their distinct advantages—CrewAI integrates well with LangChain tools for code execution, while AutoGen is noted for its customizable features and execution capabilities. Despite security concerns associated with running LLM-generated code, both frameworks have demonstrated effectiveness and utility, with developer preference often influenced by familiarity with existing tools or specific customization needs.

A code interpreter traditionally refers to a program that translates high-level programming language into machine language for execution, typically written by humans. With the advent of AI agents in 2023, code interpreters have evolved to include AI agents that translate natural language prompts into code instructions, offering capabilities beyond mere code generation. LLM-powered code interpreters, unlike traditional AI coding apps, can execute code, create files, download resources, and perform operations such as generating charts from data. Notable examples include Open Interpreter, Microsoft's Autogen agents, and ChatGPT's Code Interpreter feature, which can control computer operations via terminal or execute code in sandbox environments. In contrast, other AI tools like standard ChatGPT, Phind, Perplexity, and GitHub Copilot lack these execution capabilities, focusing instead on generating code in text format without altering system files or directories.

Many developers are creating their own AI coding agents inspired by frameworks like Open Interpreter, Autogen, and ChatGPT Code Interpreter, which have evolved to not only generate but also execute code, offering practical applications such as data analysis and visualization. Despite the advancements, developers face challenges such as ensuring security and isolation while running AI-generated code locally, often using containers like Docker, which can pose isolation risks. The user-centric approach initially favored developers, but now demands better user experiences, especially for non-technical users, who struggle with local installations and terminal controls; browser-based applications offer more intuitive interfaces and collaboration features. Scalability is another critical issue, as deploying AI agent instances for many users requires numerous isolated environments, complicating management. Additionally, there is a need to maintain long-running sessions to allow users to continue their work seamlessly. To address these challenges, platforms like E2B provide cloud runtimes for AI applications, offering secure and scalable environments that mimic local usage but with enhanced safety and usability.