Alibaba, Baidu and DarwinMind Push AI Agents From Chat to Task Completion

WeCom users can authorize their accounts through the Qianwen Office connector and then use dialogue commands to read smart spreadsheets, create or read documents, send notifications, check schedules, book meetings, or create to-dos. The web version supports the feature, with a desktop client entrance to be launched soon, Leiphone reported. Qianwen Office started public beta testing on Aug 3 and is positioned as an enterprise-level agent product. It has also integrated with Slack, and the company plans a standalone app and an international version.

Baidu announced on Aug 14 at its AI Day that its Baidu Wenku and Wangpan assistant, codenamed GenFlow, was officially named Kuku AI. The company launched Kuku AI as a standalone product across desktop, web, mini-program, and enterprise editions, Leiphone reported. Baidu described the change as a shift from a “data + AI” model to an “AI + data” model: users no longer need to locate files first and then invoke AI; they can state an objective and let the agent gather data, call tools, and finish the work. Financial research was shown as a key scenario, where the agent handles data collection, cleaning, modeling, analysis, and report generation.

Baidu said it is building a Skill ecosystem to let capabilities evolve as models improve. The company’s online library contains more than 1.8 billion professional documents, which underpins its focus on financial research, but it acknowledged that data tracing and review, especially in finance, are areas for future upgrades. The team said it does not want to build a product that retains users simply by consuming tokens. “If the model provides flour, Kuku AI should deliver bread,” a team member told Leiphone, meaning the product should hand users a complete, usable result.

Baidu also said it will not require enterprises to migrate their knowledge bases, instead directly connecting to existing storage systems. The company is targeting small and medium-sized enterprises, arguing that standardization makes it easier to lower costs across industries. Its ToB team for Wenku and Wangpan has about 100 people and has served hundreds of thousands of enterprises, according to Leiphone.

According to iiMedia Research, China’s AI agent market reached 80.4 billion yuan in 2025, up 123.2 percent year-on-year, and is expected to hit 696.8 billion yuan by 2030. Office productivity is considered a core growth area, with 59.1 percent of surveyed users saying they are willing to pay for office AI agents, Leiphone reported.

In a separate development covered by QbitAI, Hangzhou-based DarwinMind unveiled Spellcaster, a system that uses six specialized AI agents—covering rules, levels, assets, playability, simulation, and repair—to generate, playtest, and fix games. The process runs a “generate-run-check-fix” loop to address what the team calls the “playability black hole,” where code compiles and runs but the game remains unplayable or unbalanced. A user prompt such as “a bullet-hell shooter with a starfield background” can yield a playable prototype in about 15 minutes. Supported genres include platformers, tower defense, endless runners, dungeon roguelikes, and bullet-hell shooters.

DarwinMind’s core members are from Zhejiang University, Nanjing University, and Australian National University, with several co-founders serving as doctoral supervisors at Zhejiang University focusing on world models, multimodal large models, and agent systems. The company said its next phase is to use a world model to directly generate game frames and responses from player actions, removing traditional code and engine rendering as the central intermediate layer. Beta applications are open at spellcaster.world.