Harmony Harness Reframes Agent Evolution

The paper’s strongest idea is treating the harness—not just the model—as a composable optimization target. Its results suggest better orchestration and explicit security pressure can produce agents that generalize beyond the benchmark they evolved on.

• –Decomposition-based evolutionary search preserves trade-offs between functionality and security instead of collapsing them into one score.
• –Factorized operators separate general-purpose improvements from environment-specific fixes, improving transfer.
• –Dynamic workflow planning beats fixed pipelines, with extra execution steps spent selectively where they help.
• –Cross-benchmark results are especially compelling: Harmony Harness improves secure-pass performance even when evaluated on an unseen benchmark.
• –The main practical limitation is operational cost: self-evolving harnesses require repeated executions, evaluations, and careful safeguards against metric overfitting.