Recursive Harness Self-Improvement for Agent Trace Quality
July 16, 2026
Recursive Harness Self-Improvement (RHI) optimizes task-specific agent harnesses through iterative prompt-level specification updates. This method improves the quality of execution traces used for future model training without the high cost of manual scaffold updates.
HOW THIS AFFECTS YOU
●
builderYou can automate the optimization of agent scaffolds to produce better training data.
●
researcherThis enables a more efficient model-harness co-evolution cycle.