阿里字节论文：AI Agent 推理不再是性能瓶颈
New Alibaba ByteDance paper shows that AI agents can no longer be served like or…
New Alibaba ByteDance paper shows that AI agents can no longer be served like ordinary LLM requests.
阿里巴巴与字节跳动联合发表的新论文表明，AI Agent 已无法像普通 LLM 请求那样提供服务。
Because most of the performance problem now sits across tools, memory, environments, and the model together.
因为目前大多数性能问题出现在工具、记忆、环境与模型之间的交互上。
Bottleneck can also move from the LLM to embeddings, databases, sandboxes, or network transfer as the request and deployment change.
随着请求和部署方式的变化，瓶颈也可能从 LLM 转移到嵌入（embeddings）、数据库、沙箱或网络传输中。
This paper builds AgentSysBench around 10 agentic applications and finds that model inference is often no longer the main bottleneck.
该论文围绕 10 个智能体应用构建了 AgentSysBench，发现模型推理往往不再是主要瓶颈。
They find task-aware serving cuts latency by 29–40%, communication-aware placement delivers up to a 4.5× speedup, state offloading cuts memory by 4.6×, and caching removes 35.2% of redundant search calls.
研究发现，任务感知服务可将延迟降低 29%–40%，通信感知部署可带来最高 4.5 倍的加速，状态卸载可减少 4.6 倍的内存占用，而缓存则消除了 35.2% 的冗余搜索调用。
So optimizing tokens per second is no longer enough. Agent infrastructure has to schedule models, tools, memory, and communication as one workload.
因此，仅优化每秒处理的 token 数已不再足够。Agent 基础设施必须将模型、工具、记忆和通信作为统一的工作负载进行调度。
– arxiv. org/abs/2608.15127
– arxiv.org/abs/2608.15127
Title: "From LLM Inference to Agentic Workloads: Characterization and Implications for Serving Systems"
标题：《从 LLM 推理到智能体工作负载：特征分析与对服务系统的影响》
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力