Lately, there has been progress in leveraging Large Language Models (LLMs) to augment their capabilities and overcome limitations such as inability to get exact answers to questions requiring specific reasoning or dynamic reaction to recent knowledge beyond a particular context window. Researchers have built systems around LLMs that can interact with multiple underlying models optimized for different aspects of complex workflows, enabling AI agents to interact with these models to outsource other kinds of reasoning. This has led to the development of code execution applications where LLMs generate algorithms in the form of code and can be used to solve problems by running sequences of instructions. To make this work, a sandboxed environment is needed to evaluate untrusted code without causing catastrophic accidents, which is challenging for most users. Researchers have proposed two approaches to address this: one stateless API container server and another more stateful agent environment that uses a full Replit as the sandbox. The first approach has been released as a self-serve platform for experimentation, allowing users to easily customize Docker container images and integrate it with OpenAI applications. The future plans include further experimenting with new ways of augmenting LLM capabilities.