The deployment of foundation models on complex, real-world workflows—from enterprise application automation to multi-stage research pipelines—traditionally demands substantial domain expertise in prompt design, tool selection, and orchestration logic. The authors propose a principled alternative: a nested optimization framework that treats harness engineering itself as an automatable problem amenable to meta-learning.
The architecture operates across two coupled loops. The Harness Evolution Loop optimizes an agent's configuration ℋ for a specific task through three specialized components: a Worker Agent W_ℋ executes task instances, an Evaluator Agent V performs adversarial failure diagnosis and assigns performance scores, and an Evolution Agent E proposes modifications to ℋ conditioned on the trajectory of prior attempts. This creates a feedback mechanism where failures become explicit training signals for harness refinement.
Crucially, the Meta-Evolution Loop operates at a higher abstraction level, optimizing the evolution protocol itself Λ = (W_ℋ, ℋ^(0), V, E) across heterogeneous tasks. By learning a protocol Λ^(best) that generalizes across diverse domains, the framework achieves rapid convergence on entirely novel tasks—eliminating the need for human-in-the-loop harness engineering. The formalization explicitly connects this approach to meta-learning theory, enabling principled analysis of convergence properties and sample complexity.
This work represents a significant shift in the automation pipeline: from manual harness engineering to automated harness engineering to automated design of the automation itself. The implications extend beyond convenience—by learning task-agnostic evolution strategies, the framework potentially discovers fundamental principles governing effective agent configuration across domains.
