它审的是活儿，不是提示词
Cortex 读的是真实的对话记录和结果——被要求做什么、智能体做了什么、在哪里卡住或转交给了人——而不是对着一个配置文件瞎猜。
Olano Cortex
对话结束了，思考仍在继续。
Cortex reads what your agents actually did, finds the patterns, and acts on them - on a schedule, without being asked. It ships with Fully Autonomous on, so improvements land on their own - every one logged, and what a phase applies itself snapshotted with one-click undo. Rewriting an agent's own instructions, sending anything outside, and anything irreversible wait for a person in either mode. If you would rather sign off on everything, turn Fully Autonomous off and the same work queues for you to approve, defer or reject.
Olano Cortex
它会研究你的智能体刚刚过完的这一天。
大多数 AI 工具转头就忘了刚过去的一天。Cortex 会去复盘：定了什么、什么在重复、缺了什么、哪些值得再看一眼——再把这些变成更精准的记忆、可复用的技能、经过调研的想法，以及交给你审阅的改进方案。
节奏由你定——宽松、均衡或激进
自我修改一律排队等审批，绝不自动生效
Every change logged; what it applies itself, snapshotted with one-click undo
每日推理预算，事先设好上限
It runs whether or not you open the dashboard
Each phase is scheduled at a cadence you pick - relaxed, balanced or aggressive. The Think button runs the same phases on demand; it is the manual override, not the main path.
Fully autonomous, or every change signed off
One switch, and it is yours. Left on - the default - Cortex applies what it learns as it learns it: skills, sharper memory, better procedures, each logged and reversible, with the scheduled passes running behind a snapshot you can undo in one click. Three things wait for you in either mode: a rewrite of the agent's own instructions, anything sent outside your deployment, and anything that cannot be undone. Turned off, every output waits in a queue with the reason and the conversations that prompted it attached, and you approve, defer or reject each one.
它会真的写入记忆和技能
产出的不是建议。是更精准的记忆、可复用的技能和更好的流程，直接写进那个需要它们的智能体里。
节奏和预算由你定
运行频率、每日预算、跑哪些阶段，都可以按智能体分别设定，所以这个自我改进的循环花多少钱由你说了算。
阅读手册
设置、阶段、预算和自检都完整记录在 Cortex 那一章.