Exo Lab Teaches Recursive Self-Improvement

Exo’s key insight is that self-improvement needs durable history and rollback, not just editable prompts or memory.

• –Append-only event logs preserve lineage and prevent agents from repeating failed experiments
• –Sandboxed snapshots enable risky harness changes without permanently corrupting the agent
• –Agents can modify prompts, tools, memory, policies, and even clone or rebuild themselves
• –The architecture separates trusted state management from swappable executor logic
• –It remains experimental: safety depends heavily on sandbox boundaries and policy controls