How Baidu’s DaZi Upgrade Aims to Let Agents Handle Over 90% of Human Work

Agent technology has moved from impressive demos to daily workflows, exposing gaps such as incomplete task handling, unstable cross‑system execution, and limited team‑wide efficiency gains. Baidu’s DaZi addresses these gaps with a three‑pronged upgrade announced at the AIDAY event on July 10: a comprehensive personal version, the debut of an enterprise edition, and the launch of the DaZi Alliance.

Personal version upgrades focus on four dimensions—brain, hand, skills, and suite. The “brain” upgrade introduces an environment routing mechanism via the Harness engine, automatically selecting cloud sandbox, local sandbox, or chat mode based on task requirements. This reduces average task latency by 20% and improves token utilization by 25%.

Multi‑device shared memory allows users to continue a partially completed task on a different device without re‑specifying preferences, templates, or tone, ensuring seamless continuity across PC and mobile.

The “hand” upgrade improves tool stability: a cloud‑based browser on mobile (controlled via VNC) and an upcoming embedded desktop browser lower web‑operation latency by 13% and increase smoothness by 20%.

Skill ecosystem expansion adds Baidu’s internal skills (Yijing, Lingyi) and third‑party services (Meituan red‑packet, delivery) to broaden task boundaries. A fully automated skill onboarding pipeline now handles development standards, security and effectiveness evaluation, and operational rollout.

The professional suite introduces a media‑focused package for creators, integrating topic selection, content generation, short‑video scripting, storyboard creation, publishing assistance, and post‑analysis into a single conversational flow, reducing a multi‑hour workflow to under five minutes.

Enterprise edition tackles asset retention, sharing, and security. Generated assets are archived for reuse, employee methodologies become reusable skills, and organizational knowledge bases store documents and industry data. Shared resources include points, skills, and outcomes, with intelligent allocation based on usage frequency.

To simplify integration, Baidu released the industry’s first enterprise‑grade Skill access standard, bridging internal systems, third‑party SaaS, and DaZi’s skill platform, lowering engineering effort and defining clear security boundaries.

Security enhancements add robust governance (role‑based access, usage statistics), compliance auditing (recorded AI actions and data access), SLA guarantees, and safety guards that block high‑risk operations.

The DaZi Alliance opens the platform to eleven partner companies across regions and verticals, enabling them to contribute industry expertise, customer networks, and deployment services while Baidu provides the underlying AI capabilities.

Performance metrics show rapid adoption: daily user queries have risen 20‑fold since March, and DaZi topped two Agent benchmark leaderboards on May 8—scoring 93.3% and 93.2% on PinchBench (covering 23 real‑world scenarios) and 58.03 on DeepResearch Bench (focusing on retrieval, accuracy, depth, and report readability). Monthly visits grew 114.72%, reaching 1.16 million.

Baidu’s vision, expressed as the DAA (Daily Active Agents) metric, shifts focus from token consumption to the number of agents actively delivering results. The roadmap—personal upgrades, enterprise depth, and alliance breadth—aims to let agents assist in over 90% of human work, aligning with McKinsey’s estimate that 57% of U.S. work activities are automatable.