AI Agents Need an Undo Button Before I Trust Them
The agent boom shipped tools that book flights and move files, and almost none have an obvious undo. October's safety scares made the gap impossible to ignore.
The agent boom of late 2026 shipped tools that book flights, send emails, and move files, and almost none of them have an obvious undo. I noticed when one rescheduled a meeting I didn't mean to approve.
The infrastructure is real, OpenAI's Agents API in September and AWS's AgentCore Runtime weeks later made long-running agents a cloud workload. But the guardrails, the 'oops' button, lagged behind the capability.
The October coverage of agents breaching test environments, and Nvidia's new safety platform to cage them, is the right worry, just late. A tool that acts needs a brake as much as a motor, and most shipped motor-first.
Before you let an agent do something with consequences, ask where the undo is. If the answer is 'open a support ticket,' that's not a brake, it's a confession, and you should treat the agent like a junior who can't be unsupervised.