Proactive agent
Also known as: proactive AI agent, initiative-taking agent
Most AI tools are reactive: you ask, they answer. A proactive agent inverts that relationship. It watches connected data sources (email, calendar, Slack, dashboards), decides on its own what deserves your attention, and reaches out with a summary, a completed task, or a question before you ever open a chat window. The agent has a persistent process running in the background with its own schedule, not just a thread waiting for your next message.
The pattern crystallized in late 2026 as multiple major labs shipped proactive agents within weeks of each other: Anthropic's Claude Duet, OpenAI's Dots, and Meta's Muse all follow roughly the same model. Each morning brief, each unsolicited status update, each task completed while the user's phone was off is the proactive pattern in practice. The design challenge is figuring out what the agent should decide for itself versus what it should always ask about first.
The key safety consideration is approval gating. Because a proactive agent acts without being directly prompted, the stakes of a mistake are higher: it might send a message the user didn't intend, schedule the wrong meeting, or surface private information at the wrong time. Well-designed proactive agents have clear escalation paths, require explicit confirmation for irreversible actions, and let users tune what they want to be notified about and how often.