TLDR Tech / Dev
The Agent Harness Belongs Outside the Sandbox
AI agent harnesses, the control loops driving LLMs, can be architected either inside or outside the execution sandbox, each with distinct implications for security and multi-user environments. Mendral, building multi-user agents, chose to run the harness outside the sandbox to improve security by keeping credentials separate, have efficient sandbox lifecycle management, and simplify shared state.
Read the full story