Learning from Failure: RHO and SESA Advance Agent Self-Improvement

imjustnewatai · x · 2026-08-06

The author highlights two cutting-edge research approaches enabling AI agents to improve by reflecting on their past failures:

RHO (Retrospective Harness Optimization): A self-supervised method that optimizes an agent's skills, tools, and workflows without needing an external ground-truth validation set. It selects challenging tasks from past trajectories, re-solves them, and uses self-validation and pairwise self-preference to generate and select the best harness updates. A single optimization round boosted SWE-Bench Pro pass rates from 59% to 78% while keeping the underlying model fixed.
SESA: This method trains agents to judge their failures, convert them into reusable remembered skills, and generate harder future training problems. Notably, some improvement persisted even after external memory was removed, suggesting that the learned lessons had been absorbed into the model's weights.

More from coding & agent