Claude 上线 agent 执行步骤追踪（公测）
Anthropic 在 Claude 中加入 agent 追踪：记录每步耗时、token、工具调用顺序，详细模式还存输入输出，可接入现有监控工具。
Anthropic just made it possible to see exactly what your AI agent did, step by step Until now, when an agent gave a wrong answer, there was no way to check where it went wrong. It was a black box. Now Claude records every step an agent takes. How long each step took, how many tokens it used, which tools it called and in what order. Turn on the detailed mode and it also saves what the agent was asked, what it replied, and what each tool sent back. All of this flows into the tracking tools your team already uses, so there is nothing new to learn. Think of it as a replay button for your agents. Something breaks, you rewind and watch the exact step where it went off track. Separate startups were charging money for exactly this. Anthropic has now built it into Claude itself. It is in public beta, and it works across the Claude API, their managed agents and their developer tools. Agents are being trusted with real work now. Real work needs a record of what happened. This is that record.
这条是引用着另一条写的：x.com/2024518793679294464/status/211042173202906…