We present ReflexGrad, a dual-process architecture for within-episode failure recovery in LLM agents without demonstrations. When agents commit to a wrong approach early and exhaust the step budget, the post-failure trajectory contains the information to escape -- but no published architecture acts on it within a single episode. ReflexGrad routes between a fast process (TextGrad-style continuous refinement every $k{=}3$ steps) and a slow process (Reflexion-style causal diagnosis when $m{=}5$ consecutive low-progress scores fire a routing gate). A deterministic priority merge keeps the natural-language policy coherent, and each slow activation emits three observable artifacts: a reproducible trigger, a causal diagnostic, and a verified fix. On ALFWorld 134 tasks, $n{=}10$ seeds, no demonstrations, ReflexGrad lifts Qwen-3-8B from $35.1\%$ to $75.4\%$ ($+40.3$pp), beating compute-matched 1-shot LATS by $+2.7$pp ($p{\approx}0.01$), ToT by $+5.7$pp ($p{<}10^{-4}$), and Self-Refine by $+6.7$pp ($p{<}10^{-5}$); on GPT-5 the lift is $46.3{\to}88.1\%$ ($+41.8$pp). The $1.5$pp cross-model difference is within seed noise ($p{\approx}0.13$), suggesting that the routing mechanism, rather than model scale, is the primary source of the gain. Code, prompts, per-seed logs, and sensitivity sweeps are released.

翻译：我们提出ReflexGrad——一种无需示范即可实现LLM智能体片段内故障恢复的双过程架构。当智能体早期错误决策导致步骤预算耗尽时，故障后轨迹中实际蕴含逃逸信息，但现有架构均未能在单一片段内对此加以利用。ReflexGrad在快速过程（每$k{=}3$步执行TextGrad风格连续优化）与慢速过程（当$m{=}5$个连续低进度分数触发路由门时执行Reflexion风格因果诊断）之间切换路由。确定性优先级合并机制确保自然语言策略的一致性，每次慢速激活生成三个可观测产物：可复现触发器、因果诊断结果与已验证修复方案。在ALFWorld 134个任务的$n{=}10$种子实验中（零示范条件下），ReflexGrad将Qwen-3-8B从35.1%提升至75.4%（提升40.3个百分点），以$+2.7$个百分点（$p{\approx}0.01$）超越计算量匹配的单次LATS，以$+5.7$个百分点（$p{<}10^{-4}$）超越ToT，以$+6.7$个百分点（$p{<}10^{-5}$）超越Self-Refine；在GPT-5上升幅为46.3%→88.1%（提升41.8个百分点）。1.5个百分点的跨模型差异在种子噪声范围内（$p{\approx}0.13$），表明增益主要来源于路由机制而非模型规模。代码、提示词、逐种子日志及敏感性扫描已开源。