Summary
Alex Krentsel, a UC Berkeley PhD student and co-founder of Exo (backed by a16z partners Martin Casado and Anker Goel), joins the Latent Space podcast to introduce a new agent framework built around recursive self-improvement. The core thesis: rather than relying on an outer optimization loop to improve an agent, the agent itself should be able to safely inspect, edit, and rebuild its own code and policies at runtime — what Krentsel calls “collapsing the loop.”
Krentsel walks through Exo’s three-layer architecture: a stateful harness storing conversation history and state, an executive layer running policies and LLM calls, and isolated sandboxes for safe code execution. The key mechanism is that Exo mounts its own executive code inside the sandbox, enabling the agent to modify itself mid-run. A guardian process monitors each rebuild and automatically rolls back to the previous state if the new version breaks during a trial step — providing safe atomicity for self-modification.
The discussion explores why the harness layer is becoming as strategically important as model weights themselves, how token costs are driving demand for more efficient agentic architectures, and how agent state portability enables seamless migration between local and cloud execution. Krentsel draws on his prior research at Berkeley’s Sky Lab on AI-driven discovery systems, and the podcast covers how early work by OpenAI’s procedural intelligence team inspired similar thinking about agents modifying their own scaffolding.
📺 Source: Latent Space · Published August 15, 2026
🏷️ Format: Interview
