ModularRSI evolves agent harnesses via recursive self-improvement

Automated harness evolution represents a far higher-leverage frontier than manual prompt engineering, demonstrating that agent scaffolding itself can be systematically self-improved.

• –Decomposing agent scaffolds into five distinct modules prevents entangled side effects and allows targeted optimization of failure-prone routines.
• –Contrastive trajectory analysis pinpoints actual behavioral root causes by comparing matched successes and failures rather than relying on noisy single-run heuristics.
• –Using a benchmark-disjoint practice pool solves the pervasive problem of scaffold overfitting, ensuring genuine out-of-distribution generalization.
• –Demonstrated cross-model transferability proves that evolved architectural patterns capture universal agent execution primitives rather than model-specific quirks.