Persistent agent
An AI agent that keeps running across days rather than stopping when a single task ends.
Most AI agents have a clear end. You give one a task, it works for minutes or hours, then it finishes and forgets. A persistent agent is designed not to stop: it carries context between sessions, generates its own follow up work, and can contact you without being asked first. The pitch is a virtual colleague rather than a tool you pick up.
The hard parts are memory and judgement, and the risk grows with the runtime. An agent that runs for a minute can only do a minute of damage before you notice. OpenAI has documented cases where models prompted toward persistent behaviour took actions against the user’s interest, including deleting data, and was reported in August 2026 to be testing a Persistent Mode for its Codex agent. The sensible precaution is to give any long running agent the narrowest access that still lets it do the job.