技术文章

A study of 519 overseas job listings, 28 roles at major Chinese firms, and 6 projects examines FDE responsibilities, growth, costs, and scalability.

用 519 条海外公开 JD、28 条国内大厂岗位和 6 个真实项目，拆解 FDE 的职责、增长、国内落地、成本与规模化条件。

TypeSafe Jev uses Choice, Score, and Noul for constrained agent decisions while code retains control of valid actions and side effects.

TypeSafe Jev 用 Choice、Score 和 Noul 处理 Agent 的受约束决策，让模型负责模糊判断，代码继续掌握合法动作和副作用。

Anthropic and Accenture extend evaluation beyond model outputs to training, deployment, monitoring, and incident processes, with independence unresolved.

Anthropic 与 Accenture 的 embedded evaluator 计划把评测范围从模型输出扩展到训练、部署、监控和事故流程，也暴露出独立性边界。

Alibaba skill-up resumes real sessions turn by turn, letting skill evaluations test context continuity, intermediate state, and per-turn assertions.

Alibaba skill-up 的 stateful 多轮模式逐轮恢复真实 Session，让 Agent Skill Eval 能检查上下文连续性、中间状态和逐轮断言。

Claude combines Chat and Cowork in one conversation, routes capabilities automatically, and keeps task state and permissions visible.

Claude 把 Chat 与 Cowork 合并到同一会话，把能力路由交给系统，同时保留任务状态和权限确认，给 Agent 产品一个具体 UX 边界。

Gemini Live Extended Thinking separates conversation turns from background task lifecycles. Clients need interaction_status to track actual completion.

Gemini Live Extended Thinking 把对话回合与后台任务生命周期拆开，客户端需要用 interaction_status 管理真正的完成状态。