KK Blog

Menu

Archives

2026 / 09

OpenAI Agents API Migration: Audit the Managed Harness

2026-09-11

最佳实践 #OpenAI Agents API#Managed Agent Harness#Agent Migration#Agent State#Agent Security#Agent Observability#Codex Harness
Audit state, authority, recovery, observability, and cost before migrating production agents to OpenAI's managed Codex harness.
机器人访问正在变成一套强制执行栈：身份、用途与默认策略

2026-09-10

解读 #Web Bot Auth#AI爬虫#Cloudflare BotBase#robots.txt#Agent身份#Bot管理#AI治理
Web Bot Auth 只能验证签名身份。机器人访问治理还需要站点偏好、用途声明、名录分类、本地策略和可审计的边缘执行。
Web Bot Auth and AI Crawler Control: The Five-Layer Enforcement Stack

2026-09-10

解读 #Web Bot Auth#AI Crawlers#Cloudflare BotBase#robots.txt#Bot Management#Agent Identity#AI Governance
Web Bot Auth verifies a signing key. Robot access needs declared purpose, directory classification, local policy, and auditable edge enforcement.
KV Cache 变成 Agent 运行时后

2026-09-09

解读 #KV Cache#Agent运行时#AsyncReasoning#LLM推理#交互式Agent#SGLang#推理状态
KV cache 操作让预训练模型并发观察、思考和响应成为可能。机制已有实验，生产运行时仍在形成，状态治理将成为新的系统边界。
实验室 Agent 需要一份测量合同

2026-09-09

最佳实践 #实验室Agent#量子校准#GPT-5.6 Sol#Codex#测量合同#科学AI#人工接管
实验室 Agent 的能力上限由测量信号、通过条件与异常升级机制共同决定。MIT 六量子比特校准案例给出了一份可迁移的测量合同。
DseWiki 事件：只读网络为何能写入

2026-09-09

调研 #DseWiki#OpenAI Agent#Agent安全#只读网络#GET请求#沙箱#事件分析
DseWiki 事件说明，只读网络应按外部副作用定义，而非按 HTTP 方法名称定义。本文分层审计公开档案、官方承认、站长证词与研究者推断。
KV Cache as an Agent Runtime: What Changes

2026-09-09

解读 #KV Cache#Agent Runtime#AsyncReasoning#LLM Inference#Interactive Agents#SGLang#Inference State
KV-cache manipulation lets pretrained LLMs observe, reason, and respond concurrently. The mechanism works; the production runtime is still emerging.
Quantum Calibration AI Agents Need a Measurement Contract

2026-09-09

最佳实践 #AI Lab Agents#Quantum Calibration#GPT-5.6 Sol#Codex#Measurement Contract#Scientific AI#Human Oversight
Quantum calibration AI agents need every action tied to a measurable signal, a pass condition, and an escalation path. MIT's case shows the contract.
The DseWiki Incident: When Read-Only Web Wrote

2026-09-09

调研 #DseWiki#OpenAI Agents#Agent Security#Read-Only Web#GET Requests#Sandbox#Incident Analysis
The DseWiki incident shows why read-only web access must be defined by side effects, not HTTP method names. Here is what the public evidence proves.
Astra 在 ARC-AGI-3 得分 62.7% 或 99.9%：Harness 本身就是评测对象

2026-09-05

解读 #GPT-6 Astra#ARC-AGI-3#AI Agent评测#Provider Adapter#Harness Engineering#模型评测#上下文压缩
Astra 在不同 Harness 下得到 62.7% 和 99.9%。本文拆解分数语义、状态管理差异，以及速度和 token 数据的适用边界。
Previous Next