Most AI assistants wait for you to type something. A new wave of proactive AI agents does the opposite: it watches your connected apps, notices what needs attention, and reaches out first. In the past four weeks, both OpenAI and Anthropic have shipped products built on exactly this idea, and the way they handle trust, control and cost will shape how entrepreneurs and teams delegate work to software.

OpenAI introduced Dots at DevDay on September 29, 2026, and Anthropic followed on October 21 with a research preview of Claude Duet. Both give an agent its own cloud computer and browser, and both promise to keep working while you are away. This guide explains what each one does, how they differ, and what a founder or developer should do before handing real work to an always-on agent.

What Is a Proactive AI Agent?

A proactive AI agent is software that acts without waiting for a prompt. Instead of answering one question at a time, it runs continuously, checks your connected accounts on a schedule, and surfaces tasks, drafts or alerts on its own. The agent has memory of your preferences, a place to run (usually a cloud virtual machine with a browser) and a set of connected apps it can read.

According to MacRumors’ coverage of the launch, Claude Duet sends a short morning brief, checks apps every few hours, and alerts you when something needs action. It can handle simple admin work as well as larger jobs such as building a web app. Everything happens in one ongoing thread rather than a stack of separate chats, which makes it feel closer to a colleague than a chatbot.

This is the practical difference between an assistant and an agent. An assistant is a tool you pick up. An agent is a process that keeps running, which means the questions that matter shift from “how good is the answer?” to “what is it allowed to do, and who checks its work?”

Claude Duet vs OpenAI Dots: Always-On AI Agents Compared

The two launches share a design but differ in details. Anthropic describes Duet as built on Claude Opus 5.5. It is a research preview available by waitlist to US Pro and Max subscribers, on the web and a standalone iOS app, and its background checks count against your existing plan limits.

OpenAI’s Dots, as reported by The Next Web, run on GPT-6 Astra. Each dot has its own cloud computer and browser, can work on several projects at once, and connects to more than 4,000 apps through plugins. You can reach a dot from ChatGPT, Slack and Microsoft Teams, and you can watch its computer at any time. Dots are rolling out to Pro and Business Premium users in eligible markets, with the first dot included at no extra cost, and conversations with a dot do not count toward ChatGPT usage limits.

In short, Dots lean toward breadth and workplace chat integration, while Duet leans toward a single continuous thread with memory and skills carried over from your Claude account. Neither is a finished product. Both companies describe them as early, and both acknowledge mistakes are possible.

AI Agent Approvals and Safeguards: What Each Company Built In

Giving software the freedom to act raises the stakes on control. Both products answer with approval gates, though the specifics differ.

Anthropic says Duet asks for your approval before it sends a message, makes a payment, signs a document or deletes a file. It does not see your passwords or logins when given account access, and a dedicated classifier plus a safety system prompt screen each message to guard against leaks of sensitive information. Anthropic also says Duet may misread threads and cannot yet show its browser, and it has said it will share more about the safety design in the coming days.

OpenAI says background work by a dot is limited to read-only access to connected apps. Users can set custom rules to allow, block or require approval for specific actions, an auto-review step checks actions that affect accounts or share information, and a monitoring system can pause or stop a dot over safety concerns. Sensitive tasks, such as changing a password, always stay with the user. OpenAI also says specialist dots with their own identity and credentials are in enterprise pilots.

The pattern is clear: read freely, act only with permission. If you are evaluating any always-on agent, treat that boundary as the first thing to test. For a deeper look at the risks, see our guide to AI agent security rules before you deploy.

How to Use a Proactive AI Agent Safely: Practical Takeaways

Whether you join a waitlist or build your own, a few habits apply to every proactive agent.

Start read-only. Let the agent summarize and flag before you allow it to send or change anything. A week of morning briefs will show you how often it misjudges priorities.

Keep approvals on for anything that moves money, sends external messages or deletes data. These are the actions that are hard to reverse, and they are exactly the ones both vendors gate by default.

Watch the meter. Background checks consume plan limits in Duet’s case, so measure how much a week of idle monitoring really costs before you scale it across a team.

Scope access narrowly. Connect only the accounts the task needs, and review what the agent can see. Our post on AI agent governance and Workday’s agent harness shows how larger organizations think about permissions and audit trails, and our look at Google’s Gemini coworker agent covers a similar identity-based approach.

Finally, verify consequential output yourself. OpenAI states plainly that dots can still make mistakes and that consequential work should be reviewed. That advice applies to every agent on the market.

The Future of Proactive AI Agents

The direction is hard to miss. Google’s Gemini agent, Microsoft’s always-on Copilot features, OpenAI’s Dots and Anthropic’s Duet all point to the same shift: from software you operate to software that operates alongside you. Our earlier look at long-running AI agents and Microsoft Autopilot traced the start of this trend, and the last month has made it mainstream.

There are open questions. Pricing for Duet has not been published, availability is limited to certain regions and plans, and independent evidence on reliability is still thin. Company claims about safeguards are claims until outside testing supports them. The next few months will show whether proactive agents earn trust through consistent behaviour or lose it through a single visible mistake.

Conclusion: What to Take Away

Three points stand out. First, proactive AI agents run continuously and reach out first, which changes the work from prompting to supervising. Second, both Claude Duet and OpenAI Dots pair that freedom with approval gates and read-only defaults, so test those boundaries before you trust them. Third, these products are early, so start small, measure cost and accuracy, and keep a human in the loop for anything consequential.

Want more AI agent tools, articles and practical resources? Explore BigAIAgent for the latest guides. And here is a question to leave you with: if an agent could check your inbox, calendar and customers every few hours, which decision would you still insist on making yourself?