Hacker News
The Agent Harness Belongs Outside the Sandbox — Mendral's Architecture for Secure Credential Isolation — 126 Points
A Mendral engineering post (126 points, 91 comments on HN) argues that the agent harness (the loop that drives LLM tool calls) should run outside the sandbox, not inside. Key benefit: credentials stay out of the sandbox — the harness holds API keys, user tokens, and database access, while the sandbox contains only the agent's working environment. The post details three problems they solved to make the outside-harness model work. The HN thread became a practitioner discussion on agent security architecture, with multiple teams sharing their own harness isolation strategies.
↳ Follow the thread