The agent harness belongs outside the sandbox (www.mendral.com)

A recent announcement has highlighted the evolution of agent harnesses in large language model (LLM) operations, emphasizing a shift from traditional sandbox environments to executing agent loops outside these sandboxes. The significance of this move lies in enhanced security and operational efficiency, particularly for multi-user scenarios, where the loop can maintain API credentials and shared skills in a centralized database rather than within ephemeral sandbox environments. By decoupling the agent's harness from the sandbox, it facilitates better resource allocation—allowing the sandbox to be rapidly suspended when idle and quickly resumed during active commands. Key technical advancements include the implementation of a virtualized filesystem that allows agents to interface with skills and memories stored in a database, streamlining interactions within shared organizational contexts. Additionally, integration with systems like Inngest for persistent agent loops ensures that the functions can survive rolling deployments without data loss. The approach minimizes cold start delays through the use of Blaxel technology, reducing resumption time to 25 milliseconds. By keeping the agent’s operational capabilities consistent with existing LLM training paradigms while addressing the complexities of multi-user data consistency, this model aims to enhance both individual engineering productivity and collaborative efforts within organizations, marking a significant step forward in agent architecture.