KK Blog

Menu

Archives

2026 / 08

Admin Plugin：企业 Agent 控制面

2026-08-26

最佳实践 #OpenAI Admin Plugin#ChatGPT Work#Codex#Agent治理#RBAC#工作区管理#AI控制面
OpenAI Admin Plugin 把工作区管理变成权限感知的 Agent 闭环。真正的架构由身份、动作控制、审批、结果证据与审计组成。
Admin Plugin as an Enterprise Agent Control Plane

2026-08-26

最佳实践 #OpenAI Admin Plugin#ChatGPT Work#Codex#Agent Governance#RBAC#Workspace Administration#AI Control Plane
OpenAI's Admin Plugin turns workspace administration into a permission-aware agent loop. The real design is identity, action control, approval, eviden
Qwen3.8-27B 实战评测：先冻结变量

2026-08-26

最佳实践 #Qwen3.8-27B#LLM评测#量化#SGLang#KV Cache#推理强度#本地AI
一套可复用的 Qwen3.8-27B 实战评测合同：比较 tok/s 之前，先控制量化、引擎、缓存、上下文、推理强度、质量与并发。
How to Benchmark Qwen3.8-27B for Real Workloads

2026-08-26

最佳实践 #Qwen3.8-27B#LLM Benchmarking#Quantization#SGLang#KV Cache#Reasoning Effort#Local AI
A practical Qwen3.8-27B benchmark contract that controls quantization, engine, cache, context, reasoning effort, quality, and concurrency before compa
OpenAI Jalapeño 基准审计：全栈推理闭环

2026-08-26

解读 #OpenAI Jalapeño#AI推理芯片#InferenceX#AI基础设施#基准审计#能效#Agent推理
审计 OpenAI Jalapeño 首批基准：1.5 至 1.9 倍每瓦吞吐说明了什么，遗漏了什么，以及全栈推理为何才是真正的战略资产。
OpenAI Jalapeño Benchmark Audit: Full-Stack Inference

2026-08-26

解读 #OpenAI Jalapeño#AI Inference Chip#InferenceX#AI Infrastructure#Benchmark Audit#Power Efficiency#Agentic Inference
An audit of OpenAI Jalapeño's first benchmarks: what the 1.5–1.9x performance-per-watt lead proves, what it omits, and why full-stack inference is the
MTIA 300 与 MetaRoCE：两条端点设计如何处理 AI 网络瓶颈

2026-08-25

调研 #MTIA 300#MetaRoCE#AI网络#RDMA#Meta AI基础设施#HCCL#以太网#AI加速器
MTIA 300 把网络放进加速器封装，MetaRoCE 把传输智能放进端点 NIC。本文拆解两条设计线的共同目标、机制差异与证据边界。
MTIA 300 and MetaRoCE: Two Endpoint Designs for the AI Networking Bottleneck

2026-08-25

调研 #MTIA 300#MetaRoCE#AI Networking#RDMA#Meta AI Infrastructure#HCCL#Ethernet#AI Accelerators
MTIA 300 moves networking into the accelerator package; MetaRoCE moves transport control into the NIC. See how the endpoint designs differ.
怎样写出真正有效的 Agent Skill：8135 条运行记录的证据审计

2026-08-25

最佳实践 #Agent Skill#SKILL.md#程序性记忆#Coding Agent#Agent评测#上下文工程#LLM Agent
8135 次 Agent 运行显示，Skill 的主要价值是稳定动作和验证闭环。本文拆解证据边界，并给出可执行的 SKILL.md 设计检查单。
How to Write Agent Skills That Work: An Evidence Audit of 8,135 Runs

2026-08-25

最佳实践 #Agent Skills#SKILL.md#Procedural Memory#Coding Agents#Agent Evals#Context Engineering#LLM Agents
A study of 8,135 agent trials finds skills work mainly as procedural anchors. Use this evidence-based checklist to design better SKILL.md files.
Previous Next