At the 2026 8th Beijing AI Conference held at the Beijing Zhongguancun International Innovation Center from June 12th to 13th, Yin Dawei, Vice President of Baidu Intelligent Cloud, delivered a speech.

Yin Dawei highlighted that software is increasingly becoming "agentified." In simple terms, this means software that was originally designed for human use is now evolving into software for AI agents. This shift brings about three specific changes.

Firstly, capabilities are becoming skill-based. This type of software is becoming callable, orchestratable, and modular. Secondly, interfaces are becoming API-based and client-based, enabling standardization, automation, and governance. Thirdly, output content is transitioning to a format readable by agents, meaning it becomes structured, comprehensible, and executable.

He believes that while general-purpose AI agents are beginning to add value in productivity scenarios, their implementation is not immediate and faces three main categories of challenges.

The first category involves deployment issues. During deployment, the variety of platforms, such as Mac, Windows, and Linux, creates compatibility problems across multiple platforms and environments. Deployment becomes even more complex in private environments.

The second category pertains to security concerns. When using general-purpose agents like "Lobster," data security and privacy leakage become sensitive issues, posing compliance challenges for governments and enterprises.

The third category concerns permission management. Yin emphasized that for an agent like "Lobster," more permissions are not necessarily better. While extensive permissions can facilitate task completion, they also lead to uncontrollable operational risks. Conversely, overly restricted permissions can prevent certain tasks from being completed. "Therefore, managing permissions presents a significant challenge in deploying general-purpose agents."

Yin Dawei introduced that Baidu launched its general-purpose agent "DuMate" at the end of March this year, aiming to unlock productivity value while addressing these three types of challenges. DuMate possesses the following characteristics.

The first is end-cloud synergy. It features sandboxes both on the device and in the cloud. For tasks requiring 24/7 monitoring, they can be executed in the cloud, allowing operations to continue even when a user's computer is turned off. Simultaneously, for data-sensitive tasks, users can mandate that data files do not leave the local domain and are processed on the device.

The second is security and controllability. Due to its sandbox capabilities, DuMate will prompt users for secondary confirmation for risky operations, thereby enhancing security.

The third is self-evolution capability. "This is one of its most important features," he said.