可预测代理系统的Harness工程:确定性执行约束的实证研究
原标题:Harness Engineering for Predictable Agentic Systems: An Empirical Study of Deterministic Execution Constraints
AI 摘要
该研究通过实验证明,为LLM代理添加确定性执行层(有限状态控制器、强制单工具选择、输出验证、有界重试)对可复现性的影响是混合的:在四个模型-任务组合中,一个组合显著提升,两个显著降低,一个无显著影响。进一步诊断发现,自由文本规划步骤是剩余变异的主要来源;引入结构化规划(在执行前验证计划符合固定模式)后,所有组合的可复现性和确定性达到完美,任务成功率提升,但延迟影响因模型而异(Qwen变快,Gemma变慢)。研究表明,harness engineering是提升代理可靠性的有效方法,但其成本依赖模型,需实际测量。
正文节选
Harness Engineering for Predictable Agentic Systems: An Empirical Study of Deterministic Execution Constraints Abstract Large Language Model (LLM) based agents exhibit substantial run-to-run execution variance even when given identical tasks, tools, and environmental conditions — a property that is acceptable for exploratory use but operationally unacceptable in regulated domains such as finance, compliance, and auditing. We study harness engineering: wrapping an agent in a deterministic execut