Databricks 推出 Omnigent 策略:在动作发生前控制 Agent 成本与风险
Databricks 出了个 Omnigent 策略,能在 Agent 动手前拦截超支和高危操作,还能累积风险分自动要求审批,做 Agent 的可以看看。
Databricks 介绍了 Omnigent 策略机制,用来限制 Agent 的花费和危险操作。它会在动作执行前做检查,支持设置消费上限和限制工具使用。系统会跨会话累积风险分数,一旦达到阈值,下一步动作就需要人工审批。配套博客和 GitHub 仓库已公开。
How do you stop an agent from overspending or taking one risky action too many?
Omnigent policies can control both cost and behavior by checking actions before they happen. They can set spending limits, restrict tool usage, and track accumulated risk across a session so the next action can require approval once a threshold is reached.
🔗 Learn more in the blog: https://t.co/MIokD5sli1
🔗 GitHub repo for this series: https://t.co/aM3Mv8dlMC