Federico UlfoThis Latent Space episode is a systems conversation with Alex Krentsel about Exo, an AI agent harness built for recursive self-improvement. It is a complete harness in the sense OpenClaw, Pi or Hermes are — tools, tasks, integrations — with one crucial difference: Exo has full visibility into its own code and its own runtime logs, so it can improve every part of itself rather than only the parts an agent is normally handed.
Fully recursive, not just adaptive: Most agents self-improve at the edges — updating memory, writing a new skill. Exo can operate on any aspect of itself: prompts, memory, tooling, and the harness policy itself. It can also clone itself and manage a lineage of those clones.
The one thing it cannot touch: An immutable event log holds the canonical history of what has already been tried. Leaving that outside the agent's reach is what stops recursive self-editing from walking into a loop it can no longer see.
Safety through architecture: Krentsel's argument is that a harness is a place you can enforce properties structurally rather than bolt constraints on afterwards — "in harness space, we actually have an opportunity to enforce certain properties by the architecture of the system that we design."
Where the stack is moving: The episode frames a shift from model weights to agent harnesses as the layer where progress now happens, and covers the decomposition of Exo into Executor, Exo Harness and Sandbox, transferable agents, sub-agents and secret management, interruptible work, and normalization protocols like ACP.
Evaluation and reward hacking: How to evaluate a system that rewrites itself — and how to keep it from optimizing the scoreboard instead of the task.
A production result: A real deployment where recursive optimization drove costs down 96%.
Founder, Engineer
New York City
