评测

覆盖 2026-09-11 至 09-16 的 AI Agent 工程阅读日志，保留本期重点、工程意义、优先阅读位置与一手来源。

AI agent engineering readings for September 11 to 16, 2026, with key findings, engineering implications, suggested sections, and primary sources.

Use Kitaru Replay, Evaluator, and Regression Suite to preserve production failures as repeatable agent tests, with tool policies and CI gates.

结合 Kitaru 的 Replay、Evaluator 和 Regression Suite，说明怎样把线上失败固定成可重放的 Agent 回归测试，并用工具策略与评测结果接入 CI。

拆解 AWS 如何用确定性工作流分开 Agent 提案、规则验证、人工审批和真实执行，让模型负责候选方案，代码负责边界与幂等。

How AWS separates agent proposals, rule checks, human approval, and execution, with code enforcing boundaries and idempotency.

The Anthropic incident review shows why automated monitors and LLM graders need coverage, independence, and tests for their own failure modes.

从 Anthropic Agentic Search 事故复盘看自动监控和 LLM grader 的漏检问题，说明 Agent Eval 也需要覆盖率、独立性和失效模式测试。

AI agent engineering readings for September 4 to 10, 2026, with key findings, engineering implications, suggested sections, and primary sources.

覆盖 2026-09-04 至 09-10 的 AI Agent 工程阅读日志，保留本期重点、工程意义、优先阅读位置与一手来源。

OpenAI organization-level data suggests that more coding-agent use shifts bottlenecks to task decomposition, concurrency, harnesses, evaluation, and selection.

OpenAI 的组织级数据说明 Coding Agent 使用规模上升后，瓶颈会转向任务拆分、并发调度、Harness、Eval 和结果选择。