Artificial-intelligence surrogates can support second-by-second thermal-hydraulic forecasting, but models selected and frozen offline may become condition-locked once deployed outside their pretraining envelope. This study develops a guarded continual-adaptation framework for experimental thermal-hydraulic loop data in which role-separated agents - Monitor, Diagnosis, Adaptation, Safety-Auditor, and Orchestrator - diagnose error signatures, prioritize candidate model families, and review promotions, while deterministic champion-challenger gates and background shadow learning retain final authority over model replacement. Seven surrogate families were screened by blocked three-fold cross-validation, and a temporal Fourier neural operator was selected as the initial champion for 60-s-history-to-10-s-trajectory forecasting on two held-out transients, with three seeds per adaptive mode. Static deployment gave a channel-averaged MAE of 7.06 and a 56.8% warning-exceedance ratio; rule-based adaptation reduced MAE to 6.54, whereas shadow refresh alone remained close to Static. The MA-Full mode, in which the role-separated multi-agent council reviews every evaluated stream step, achieved the lowest mean error, 5.72, and 35.8% exceedance, corresponding to a 19.0% improvement over Static. Paired bootstrap intervals against Static excluded zero, although intervals among adaptive modes overlapped and the six paired units limit broad statistical claims. Validated promotions from the neural operator to Transformer and graph neural network indicate that logged, gate-controlled adaptation can support auditable surrogate evolution while deterministic gates retain deployment authority.

翻译：人工智能代理可支持秒级热工水力预测，但离线训练并冻结的模型一旦部署到预训练范围之外，可能陷入状态锁定。本研究针对实验热工水力回路数据，提出一种受保护的持续自适应框架：通过分工明确的监测、诊断、自适应、安全审计和编排智能体，诊断误差特征、优先排序候选模型族并审查模型升级，同时由确定性冠军-挑战者门控机制和后台影子学习保留模型替换的最终决策权。采用分块三折交叉验证筛选七种代理模型族，选定时域傅里叶神经算子作为初始冠军模型，在三个保留工况瞬态上执行60秒历史到10秒轨迹预测，每种自适应模式重复三次实验。静态部署的通道平均MAE为7.06，预警超限比56.8%；基于规则的自适应将MAE降至6.54，而仅使用影子刷新模式的结果接近静态模式。完整多智能体模式（MA-Full）中，分工明确的多智能体委员会审查每个评估流步骤，取得最低平均误差5.72和35.8%的超限比，较静态模式提升19.0%。与静态模式配对的bootstrap置信区间排除零值，但各自适应模式间存在区间重叠且六组配对单元限制统计推断范围。从神经算子到Transformer和图神经网络的验证性模型升级表明，采用日志记录的、门控控制的自适应可支持可审计的代理模型演化，而确定性门控机制保留部署决策权。