评测

5 条 Agent 工程更新，重点看 Tool 与 Skill 上下文预算、context doctor、持久调度、Tool Loop 预算和 fail-closed 审批。

Five updates on tool and skill context budgets, context doctor, persistent scheduling, tool-loop limits, and fail-closed approvals.

Use prompt audits, instruction layers, and migration evals to retain real requirements, reassess model workarounds, and enforce hard limits in code.

拆解 Prompt Audit、指令分层和迁移 Eval：模型能力变化后，哪些规则应保留，哪些模型补丁要重新验收，哪些限制必须下沉到 Hook 或策略层。

Seven updates on token efficiency, Codex networking, Copilot Memory, local agent server security, self-practice, skills, and reproducibility.

7 条 AI Agent 工程更新，覆盖长任务 Token 优化、Codex 网络策略、Copilot Memory、本地 Agent Server 安全和 Self-Practice。

Five updates on proactive safety intervention timing, state revision, TeamHarness, enterprise MCP permissions, and tool-result handling.

5 条 AI Agent 工程更新，重点看主动安全介入时机、状态修正、TeamHarness、企业 MCP 权限边界和 Tool 结果处理。

Anthropic's book market experiment locates most efficiency loss in preference representation. A look at model choice, preference evals, and market rules.

Anthropic 的 Agent 市场实验显示，主要效率损失发生在偏好表示和匹配阶段。本文拆解模型选择、偏好 Eval 与市场规则。

GitHub Security Lab Fuzzing Taskflow uses coverage gaps to guide model decisions while deterministic tools execute, measure, and persist results.

拆解 GitHub Security Lab 的 Fuzzing Taskflow，模型根据 coverage 缺口决定下一轮，确定性工具继续负责执行、测量与持久化。