文章
一个生产级 Agent Runtime 还需要状态、审批、工具、事件与可观测性。
Agent 的难点不在于让模型多调用一次工具，而在于让执行过程可控、可评测、可恢复。
Demo 里的 Agent 可以是一个循环：模型规划，调用工具，观察结果，再继续规划。但进入真实系统之后，问题会变成权限、状态、审计、幂等、超时、回滚和人工确认。
一个更可靠的 Agent Runtime 至少要管理任务状态、工具注册、参数校验、审批节点、事件日志和观测指标。模型只是其中一个决策组件。
如果一个 Agent 执行失败后无法解释它做过什么，也无法安全恢复，那么它就还不能算生产级系统。
In a demo, an Agent can be a loop: the model plans, calls a tool, observes the result, and continues.
In a real system, the problem becomes permissions, state, audit, idempotency, timeout, rollback, and human confirmation.
A reliable Agent Runtime needs to manage task state, tool registration, parameter validation, approval nodes, event logs, and observability metrics.
The model is only one decision component inside the system.
If an Agent cannot explain what it did after failure and cannot recover safely, it is not ready to be called production-grade.