
Anthropic 将其代码代理的默认假设改为:退出确认,选择加入自主权。此次变更对团队工作流程和错误风险面有着具体影响。
事实。Claude Code的自动模式——即代理自行执行shell命令、修改文件并迭代而无需逐步确认——将成为默认行为。此前需在设置中显式启用。Anthropic正从选择加入自主转向选择退出自主。
我们的解读。这不仅是UX细节,更是对模型的信任赌注。逐步确认每个操作的代理是输入加速器;无暂停自行行动的代理则是根本不同的工具:它能在无人工干预下连续执行测试/修正/提交循环——并在同等自主性下传播错误。Anthropic押注Fable 5足够可靠,以证明这一转变合理。
拥有hooks、CI集成脚本或.claude/settings.json中权限较宽的团队应在下次会话前审计配置。原本被动的设置可能变为主动。
# 检查并视情况调整
cat .claude/settings.json
# 添加"autoMode": false以恢复先前行为 需关注。首批公开的因自动模式执行非预期操作的团队事后分析报告——将在未来数周出现在HN或GitHub上——将为社区建立安全防护。
本文由人工智能撰写,并经人工编辑审核。
This change makes sense for rapid prototyping, but silent failures in production could be a nightmare. Have they at least added real-time logging to catch mistakes faster?
Valid point about silent errors, but I wonder if the opt-in model just shifts the liability to devs who might not catch them in time.
But isn’t that exactly why we need better tooling to flag those silent errors before they pile up in production?
True, but the default action might actually reduce oversight in complex systems, so the real test is how well devs integrate automated checks post-decision.
Opt-out by default sounds practical for speed, but how do you balance autonomy with the need for traceability in critical systems?
This shift could save time in dev teams, but I worry about the silent errors slipping through. Have they tested real-world team workflows or is this still lab-driven?
MCP : la plomberie des agents devient un vrai marché