Foundation model-based multi-agent systems are moving into production, yet work on runtime reliability and serving efficiency is still fragmented across diagnostics and infrastructure. This review synthesizes 137 references, 45 systems, and 8 benchmarks through a unified runtime-stack perspective. We organize diagnostics into failure localization, observability, safety, and benchmarks, and infrastructure into memory management, prefix-aware scheduling, cache lifecycle, and power management. The central synthesis is that both layers act on overlapping runtime state but optimize different objectives. Diagnostics benefits from retaining traces, context, and replayable state, whereas serving infrastructure reclaims or compresses state to satisfy memory, latency, and throughput budgets. We term this interaction the diagnostics–infrastructure retention tension. We further identify three workload characteristics—sparse activation, invocation distance, and inter-agent memory sharing—that shape agent-aware serving, and use explicit evidence tiers to separate full-text-verified results from abstract-level or contextual evidence. Rather than ranking heterogeneous systems, we derive testable co-design directions: retention-aware eviction, failure-aware scheduling, budgeted observability, and infrastructure-triggered diagnostics. This framing positions runtime state retention as a shared systems resource and clarifies where current multi-agent diagnostics and serving research remain disconnected.

在现代集合论中，奇数集与偶数集可以建立一一映射，被认为基数相等；但从几何直观来看，该结论与直观感受存在差异。本文借助数轴，从几何角度分析奇数、偶数集合的规模关系，进一步拓展到各类数集对比，并讨论实数在数轴上的分布规律，以区间长度衡量点集的几何规模。

Excessive nitrogen (N) fertilizer use has led to serious environmental issues, including N2O emissions and nitrate accumulation in intensive farmland. Manure application and wastewater irrigation are increasingly promoted as sustainable alternatives to reduce N inputs and conserve irrigation water, yet their impacts on soil N transformation mechanisms in subsoils remain unclear. Here, 0–100 cm soil samples were collected from a 5-year field experiment with three treatments: pig manure application and pig wastewater irrigation (PM+PW), pig wastewater irrigation (PW), and deep groundwater irrigation (CK). The N cycling community was investigated by quantifying genes involved in N fixation, assimilation, ammonification, nitrification, and denitrification. The PM+PW and PW treatments significantly elevated the relative abundances of AOB amoA, nirK, nosZ genes, and potential denitrification activity, but suppressed nirS gene abundance. These treatments also decreased the ratio of (nirK + nirS)/nosZ, indicating a potential effect in reducing N2O emissions. Furthermore, PM+PW and PW treatments increased NO3−-N accumulation above 50 mg·kg-1 in 0–10 cm topsoil, while reducing the NO3−-N content by over 70% in 25–100 cm soil depth. Hierarchical cluster analysis revealed the divergent distribution of N cycling genes and bacterial communities between surface (0–25 cm) and deep (25–100 cm) soil layers. Partial least squares path modeling further showed that bacterial abundance and soil properties exerted direct effects, while soil depth exerted an indirect effect on the distribution of N cycling genes. This study provided a potential approach to alleviate agricultural environmental hazards by regulating the soil N cycling microbial communities and their functions.