Exo agent harness draws buzz: full code exposed to models for recursive self-improvement

omarsar0 · x · 2026-09-03

Martin Casado highlighted Exo's results as the best price/performance, and Omar Sanseviero recommended it. Exo is an open-source agent harness built for recursive self-improvement, aligned with the bitter lesson by exposing the entire harness—running code and logs, not just prompts—to the model, with dynamic upgrades. It splits an agent into three layers: an immutable durable event log, a freely modifiable executor, and a snapshot/rewind sandbox. A free 6-lab hands-on course on DAIR Academy covers model bindings, giving agents a shell, reading event logs, forking conversations, and a capstone experiment.

Original post →

More from coding & agent

coding & agent channel →