HarnessEvolve: Studying from Reference Trajectories for Dependable Agent Self-Evolution

arXiv:2609.00829v1 Announce Sort: cross
Summary: Self-evolving brokers advance towards autonomy by optimizing their harness—prompts, abilities, instruments, and execution logic—based on environmental suggestions. This paradigm, nevertheless, is hampered by three challenges: textit{credit score project failure}, the place terminal success/failure suggestions makes it ambiguous which step induced the error; textit{shortcut studying}, the place brokers memorize task-specific patterns moderately than purchase generalizable capabilities; and textit{catastrophic forgetting}, the place unguarded updates degrade beforehand acquired competence. On this paper, we introduce HarnessEvolve, a self-evolving framework that learns from reference trajectories to realize dependable agent self-evolution. HarnessEvolve decouples the execution agent from the evolutionary pipeline, assigning execution, analysis, optimization, and gating to impartial agent modules, enabling generalizable and secure harness enhancements. Particularly, HarnessEvolve overcomes credit score project failure by producing reference trajectories (execution paths produced when given the ground-truth solutions) and aligning failed executions towards them to extract error alerts, that are clustered to disclose systematic failure patterns. To stop shortcut studying and catastrophic forgetting, candidate harness updates should go two gates: a top quality gate that filters information leakage and immediate bloat, and a efficiency gate that accepts every replace if it improves on the present batch with out degrading current batches, with epoch-end validation on a held-out set choosing the best-performing accepted agent snapshot. We conduct in depth experiments on a number of benchmarks spanning open-domain and enterprise situations, utilizing completely different fashions and agent frameworks. Outcomes display that HarnessEvolve constantly outperforms state-of-the-art baselines throughout all benchmarks and settings, confirming reliability throughout activity domains.



Source link