HELIX Enables Co-evolution for Recursive AI Self-Improvement.

The advancement of AI agent capabilities has largely focused on improving the underlying models. However, an interactive agent's performance is profoundly influenced by its runtime harness, which manages context, tools, control flow, and stopping conditions. This strong coupling suggests that true recursive self-improvement requires the co-evolution of both the model and its harness. The process involves building harnesses for a stable model, updating the model using verified successful trajectories, and then rebuilding harnesses as the model's capabilities evolve. Realizing this co-evolutionary loop demands a controlled method for evolving harnesses while maintaining the identity and effect of interventions. HELIX provides a source-traceable substrate for this harness evolution. It systematically decomposes agent systems into typed ports, reusable atoms, recipes, product shells, and runtime policies, making all interventions explicit and auditable. This framework retains trajectories, test outcomes, and provenance, which are crucial for learning. Harness evolution within HELIX serves a dual purpose: it directly improves the execution of a fixed model and simultaneously generates matched successes, regressions, near misses, and alternative solutions. This rich dataset then becomes valuable learning signal for subsequent model improvements. An evaluation round on code repair demonstrated that a portfolio of 65 candidate harnesses discovered one that improved task coverage by 4.0% over a baseline, while the full portfolio exposed up to 58.0% more verified coverage through complementary sibling behaviors. The system also yielded hundreds of verified records for supervised fine-tuning, critic, filter, and preference learning, illustrating how harnesses, models, and data form a continuous feedback system for recursive self-improvement.