Baidu Dazi Adds Cross-Device Handoff with Cloud Browser Execution for Continuous Agent Tasks

Forum topic · 小凯 · 2026-07-27

Summary

Baidu Dazi, an agent product from Baidu AI Cloud, has been updated to let tasks started on a desktop continue on a smartphone. The handoff transfers not only chat history but also task context, execution progress, and long-term memory. The desktop client now includes an embedded browser, while mobile devices can hand web tasks to a cloud browser that keeps running. Users can initiate research on a PC, check progress and add requirements on the go, then resume and download reports back at their desk. For sensitive actions like logins, authorizations, and form submissions, the human can take over for confirmation rather than granting blanket permissions. Baidu reports vendor-side metrics including 20% lower average task time, 25% higher token utilization, 50% lower first-token latency, and up to 75% lower credit consumption, though these lack independent benchmarks. The post also examines the underlying routing logic that picks models, execution environments, and sandboxing by task difficulty, security, and cost, and discusses open questions around permission models, audit logs, and credential cleanup before enterprise adoption.

A recent update to Baidu Dazi (百度搭子), Baidu AI Cloud's agent product, addresses a practical problem in agent products: the user walks away from the computer before the task is done. According to Baidu AI Cloud, desktop and mobile can now work in relay mode — the handoff covers not just chat history but also task context, execution progress, and long-term memory. The desktop client gains an embedded browser, and on mobile, a cloud browser can continue executing web tasks.

Use case

A user can start competitive research on a PC, check progress and add requirements from a phone while out, then return to the desk to continue editing and download the report. The browser can open multiple sites in sequence — searching product information, organizing reviews, comparing prices and features — and consolidate results into a table or document. For critical actions such as logins, authorizations, and submissions, the user can step in to confirm, rather than handing over all permissions at once.

Cross-device + browser execution, together

The design combines two capabilities that are usually separate:

Cross-device sync alone just moves the conversation to another screen.
Browser automation alone tends to stall on logins, captchas, and manual authorizations.

Baidu Dazi tries to chain task context, web execution, and human takeover into one workflow: tasks keep running, but key checkpoints stay with the human.

Vendor-reported metrics

Official figures include:

Average task duration down 20%
Token utilization up 25%
First-token latency down 50%
100% completion rate on simple tasks; 94% high-delivery rate on complex tasks
Credit consumption reduced by up to 75%

These should be treated as vendor-reported numbers, not independent benchmarks — no sample size, task distribution, or control group is disclosed, so they cannot be used to rank against other agents.

Routing logic

More interesting than the metrics is the routing behind the product. Baidu Dazi says it automatically decides which model to use, whether to run in the cloud or locally, and whether to enter a sandbox, based on task difficulty, response speed, data security, and execution cost. Skills no longer require users to remember tool names; the system composes capabilities like research and office collaboration from the stated goal. For users, the entry point shifts from "pick a model, pick a tool" to "describe the goal"; for engineering teams, complexity moves into routing, permissions, and failure recovery.

Risks

Browser agents can touch login sessions, internal files, and submittable forms; cross-device long-term memory enlarges the data boundary. The official writeup mentions user confirmation for critical steps but does not publish a complete permission model, audit logs, or exception-recovery details. Before enterprise adoption, open questions remain: which pages can be accessed, which actions require double confirmation, and how credentials are cleared after task interruption.

Implications for AI coding

The lesson for AI coding is direct: a mature coding agent shouldn't only generate code in a terminal. It needs to view tasks across devices, look up documentation and file tickets in a browser, hand long tasks to background execution, and return critical actions to the human. Baidu Dazi has shipped this "continuous task" experience in office and web scenarios first — the question is whether it can back the promised continuity with stable permissions and auditing to make it a dependable workflow.