Anthropic 工程师讲 agent 循环：Jev 分离模型判断与确定性逻辑

kozh ./ @mr_kozh原推发布于 2026.10.25

在 X 上打开原推

Anthropic engineer: “Most people keep prompting Opus 5.5 after every failure. We build systems that preserve the goal, verify the result, remember the mistake and choose the next action themselves.” that is the difference between using an AI model and building an actual agent. most developers still work like this: prompt → wait → inspect the output → find the mistake → write another prompt → restart when the context breaks the human remains trapped inside every loop. Anthropic’s approach moves that loop into the system: Opus 5.5 → persistent goal → sandboxed tools → verification → failure memory → next action the agent reads its current state. decides what must happen next. executes the action. checks what actually changed. and if the result fails, it records why before trying a different approach. in this 30-minute workshop, she shows how to build agents that: preserve context across long tasks verify their own work recover after failed tool calls turn execution logs into reusable skills continue working without another human prompt Opus provides the intelligence. the harness controls the environment. loops keep the work moving. JEV separates model judgment from deterministic logic. verification decides whether the result is allowed to survive. this is what most people miss: the future of agents isn’t one perfect prompt. it’s a system that can generate the next prompt from its goal, state, evidence and previous failures. this free workshop beats most $1,000–$2,000 agent engineering courses. bookmark it and watch it today. then read the complete Opus 5.5 loop + JEV engineering guide below ↓

这条是引用着另一条写的：x.com/mr_kozh/status/2076677019463782910