Agent harnesses need three layers: executive, harness, sandbox

Exo introduces a three-layer agent architecture that decomposes agents into executive, harness, and sandbox layers. The executive is fully stateless and contains all policy decisions, the harness holds protected state like secrets and conversation history, and the sandbox provides isolated execution environments.

This separation allows the executive to safely edit its own runtime code because state and secrets remain protected in the harness layer. The architecture includes automatic rollback mechanisms if self-modifications break functionality.

The key claim is that recursive self-improvement is now practical because the harness is written in code, which LLMs can generate and modify. Previous self-improvement required changing model weights, which was impossible for LLMs to do directly.

Most teams will miss that this architecture solves the secret management problem for agents. The tradeoff is increased architectural complexity and the unsolved problem of specifying evaluation criteria for self-modification to prevent reward hacking.