评测

本期收录 GitHub Copilot 本地 Sandbox、Agensh、CliffCompaction、SWE-Serve，以及 Copilot OpenTelemetry。

GitHub Copilot local sandboxing, Agensh, CliffCompaction, SWE-Serve, and enterprise OpenTelemetry support in the Copilot app.

Claude Opus 5.5 agent API compatibility, RRSI controls for harness overfitting, Harness-Zero behavior distillation, and Strands harness defaults.

本期收录 Claude Opus 5.5 的 Agent API 兼容变化、RRSI 的 Harness 过拟合控制、Harness-Zero 的行为蒸馏，以及 Strands harness 的默认策略。

This Minecraft agent uses Astra for slow planning, JEV for frequent action selection, and Mineflayer for execution, with logs and scripts to verify runs.

这个 Minecraft Agent 让 Astra 负责慢规划、JEV 负责高频动作选择、Mineflayer 负责确定性执行，并用日志和验证脚本保留运行证据。

How task dependencies shape multi-agent results, with SAIGE experiments, ablations, and OpenAI guidance for deciding when to delegate.

Rethinking Multi-Agent Collaboration 说明多 Agent 的收益取决于任务依赖结构。本文结合 SAIGE 主实验、消融实验和 OpenAI 多 Agent 工程建议，整理可执行的委派判断。

Android Bench 2.0 keeps pass rate as a hard gate and adds completion rate to diagnose progress and unfinished acceptance requirements.

Android Bench 2.0 保留 Pass rate 作为硬门禁，再用 Completion rate 解释长任务已经完成多少，以及问题卡在哪一类验收。

AI agent engineering readings for September 18 to 19, 2026, with key findings, engineering implications, suggested sections, and primary sources.

覆盖 2026-09-18 至 09-19 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。