Baidu Dazi Adds Cross-Device Handoff and Built-In Desktop Browser for Continuous Complex Tasks

Key Highlights

Baidu "Dazi" unveiled several major upgrades at a recent AI Day, the most notable being interconnection between PC and phone: task context and execution progress sync in real time, so users can seamlessly hand off complex work across devices. Combined with the official launch of a built-in desktop browser, this agent now has the ability to execute long-chain tasks continuously across devices and across applications. Put simply, it turns "started on the phone, finished on the PC" into a system-level capability rather than something the user manually copies and pastes. The positioning matters because most knowledge workers no longer live inside a single screen; they start a thought on a commute and need it executed at a desk, and treating that as a first-class handoff is what separates a true agent from a chatbot. With these changes, Baidu is explicitly competing on continuity of work rather than on raw model size. The upgrade also reframes Baidu's agent as infrastructure rather than an app, implying the company wants Dazi to sit underneath daily work rather than beside it, and to be judged by whether the work actually gets finished rather than by how clever the chat feels.

Capabilities and What Happened

This upgrade centers on three points. First, cross-device interconnection: once PC and phone are logged into the same account, the agent syncs the current task's state, intermediate artifacts, and follow-up plan, so a research task half-done on the phone during a commute can be resumed with one click on the office computer. Second, the built-in desktop browser can automatically open multiple web pages to perform research and file downloads, essentially turning the browser into a "hand" the agent can directly operate. Third, cloud-based remote control on the phone side lets users issue commands through the cloud to complete operations even when they are not physically near the device. Taken together, these three capabilities form a loop in which the agent can perceive, act, and persist state across the boundary of a single machine, which is the minimum bar for anything that claims to handle "complex tasks" rather than single-shot answers. Crucially, the handoff is stateful: the agent does not merely pass a link but the entire working memory, so the desktop pick-up continues mid-thought rather than restarting from zero. That statefulness is what makes a multi-device task feel like one continuous effort instead of several disconnected retries stitched together by the user.

Technical Details

Underpinning the experience is an "intelligent router" mechanism: the system automatically matches execution modes based on task type, dynamically scheduling between lightweight Q&A and heavy multi-step tasks, thereby cutting average task time by 20% and raising token utilization by 25%. On delivery quality, simple-task completion reaches 100%, the high-delivery rate for complex tasks is 94%, and point consumption can drop by up to 75%. This means that within the same budget, users can complete more and more complex tasks, with a significant drop in unit cost. The router is the quiet hero here: by recognizing whether a request is a quick factual lookup or a long browser-driven workflow, it avoids spending premium tokens on trivial turns and reserves capacity for the steps that actually need it, which is why both time and token efficiency move in the same direction instead of trading off. Because the router sees the whole session, it can also learn from completed tasks and pre-warm the next step, smoothing perceived latency further and avoiding redundant token spend on steps the user has already approved. Over thousands of runs, those small savings compound into the published 20% time and 25% token improvements rather than appearing all at once.

Comparison With Competitors

Compared with traditional assistants that stay trapped in a chat box, Baidu Dazi's differentiator is that it "actually acts"—the built-in browser and cross-device handoff move it from answering questions to executing tasks. Against similar desktop agents, its account-level dual-device sync lowers the friction of fragmented usage, while the intelligent router's balance of cost and quality reflects a maturity of engineering deployment rather than simply piling on features. Many rivals can open a browser or fill a form; far fewer can preserve the full task graph across a device switch without the user re-explaining intent, and that continuity is precisely what makes the result feel like a colleague rather than a tool. The benchmark numbers on cost and completion rate suggest this is being measured in production, not just demoed on stage. In a market where many agents demo well but fail on long tasks, Dazi's published completion and cost numbers set a bar rivals will be pressured to match publicly rather than only claim in marketing. The account-level sync in particular is hard to fake in a demo and easy to feel in daily use, which is why continuity may become the real differentiator.

Industry Impact and Use Cases

To put it bluntly, this kind of cross-device continuous-execution agent is best suited to knowledge-work scenarios where "the person is always moving but the workload is heavy": market research, competitive analysis, document compilation, and cross-platform transfer can all flow seamlessly between devices. For enterprise users, it means more repetitive long-chain work can be handed to the agent, freeing humans for decision-making. As terminal forms diversify, embedding agents into phones, PCs, and even cars is moving from a selling point to a standard feature. The bigger implication is organizational: when an agent can carry a task across the day and across devices, the unit of work stops being "a session" and becomes "an outcome," and that shift is what lets teams measure AI by delivered results instead of by minutes stared at a chat window. If the pattern holds, "agent continuity" could become a checklist item in enterprise procurement, much like uptime SLAs are today, and the point-consumption cuts make the business case easy to defend to finance. The deeper shift is cultural: when the device boundary stops mattering, teams stop scheduling work around where they will be and start scheduling it around when it is due.