CITIC Construction Investment: Major agent products are being intensively implemented, and the AI industry trend is improving

The Zhitong Finance App learned that CITIC Construction Investment released a research report saying that recently major manufacturers have been intensively updating agent products. Among overseas manufacturers, Microsoft strives to build an agent network through local+cloud collaboration; Google builds a 2C Agent 3P strategy based on the existing ecosystem; and Athropic released the Claude 4 model to assist in agent construction. Domestic manufacturers are also following suit. Kingdee International (00268) built the Sky Agent Platform 2.0 and five major smart devices to help enterprises manage AI; Kunlun Wanwei (300418.SZ) brings an efficient office experience through Tiangong Super Smart. As agent products are implemented at an accelerated pace, the AI industry continues to improve, and the pace of commercialization is expected to accelerate.

Investment advice: Agent product intensive updates. 1) It benefits software companies with data, customers, and scenarios. AI products are expected to drive the company's ARPU increase and project unit price increase; 2) Increased demand for model privatization will benefit all-in-one computers, hyperconvergence, and B-side service outsourcing companies.

CITIC Construction Investment's main views are as follows:

Agents from various manufacturers were intensively updated this week, starting with GM and B-side to accelerate the progress of the industry:

Microsoft: Local+cloud collaboration to build a complete agent network

On May 19, Microsoft Build 2025 was held, and CEO Nadella gave a keynote speech on the era of intelligent agents (Agentic AI). Specifically, Microsoft wants to build an agent network where target accomplishments proactively initiate tasks, make independent decisions, coordinate with other artificial intelligence systems, and complete complex workflows with minimal human supervision. As far as the current layout is concerned, Azure AI Foundry is a cloud-based training center for models, where developers can freely combine AI capabilities such as models, functions, and tool scheduling to build required AI applications; Windows AI Foundry supports AI models locally deployed on the device side and can run AI models, tools, and agents without an internet connection; while the Copilot Studio platform enables all users to build their own intelligence and supports multi-agent collaboration; in addition, Microsoft also announced that it will join the MCP Steering Committee. It also launched a new open source project, NLWeb, which positions the HTML protocol in the smart networking era. It aims to simplify the development of natural language interactive interfaces for websites and easily transform any website into an intelligent application platform with AI functions.

Google: Building a 2C Agent 3P strategy based on the existing ecosystem.

The Google I/O conference was held on May 20. In addition to the model update and the full launch of the AI search engine AI Mode, Google also showed its own layout in terms of agents. Among them, Gemini is fully transformed into an agent, and the Agent Mode function supports users to require AI to complete tasks on their own; Chrome will also be connected to Gemini to allow users to perform AI and agent operations in the browser; Project Mariner has been further upgraded from the early prototype, and currently supports up to ten different tasks at the same time. Furthermore, at the press conference, Google clarified its “3P” strategy of TOC Agent — Personal (personalization), Proactive (predictive), and Powerful (ability), which means that Google will rely on Gmail, Docs, Maps, and private domain memories of Gemini and Search based on large model capabilities, so as to achieve things that users may need to be reminded of and may need to do in the future to improve the overall intelligent experience.

Athropic: Claude 4 Achieves the World's Strongest Programming Model to Assist Agent Construction

On May 23, Anthropic officially launched the Claude 4 series big model. Early models include Claude Opus 4 and Claude Sonnet 4, which have achieved better performance in code generation, advanced inference, and AI agents. Specifically, as the strongest coding model in the world, Claude Opus 4 leads the SW-bench (72.5%) and terminal-bench (43.2%) benchmarks, and can work continuously for several hours. Its ability to solve complex problems provides stronger support for agent development. GitHub has now introduced Claude Sonnet 4 as a new coding agent model to GitHub Copilot. At the same time, Anthropic officially opened Claude Code based on a new model, which can be integrated into terminals and commonly used IDEs as a programming agent, significantly improving developer efficiency.

Kingdee International: Sky Agent Platform 2.0+ Five Agents Help Enterprise AI Management

On May 20, the Kingdee Cloud Sky AI Summit was held to release 2.0 of the five major agents and the Sky Agent platform. Among them, Kingdee Cloud Home acts as an AI portal, and various agents can be called through dialogue to achieve multi-agent collaboration.

1) Golden Key Financial Report: Provides users with financial report interpretation and corporate financial analysis report writing services. Functions include financial report analysis, financial report comparison, index comparison, financial report writing, etc.

2) ChatBI: Queries data, provides individual visual charts, and provides cause analysis and suggestions.

3) Recruitment intelligence: The employment department puts forward recruitment requirements, and the agent generates detailed descriptions and automatically publishes JD; at the same time, the smart body also automatically collects resumes, matches requirements, and arranges AI interviews. The recruitment process was shortened by more than 80%, and the personnel matching rate was increased from 40% to 90%.

4) Travel intelligence: Multiple agents collaborate to plan itineraries, complete tickets and hotel reservations, and provide travel recommendations based on personal preferences. At the same time, Travel Intelligence provides corporate payment codes exclusive to the itinerary, so employees do not need to be reimbursed to ensure that expenses are in compliance.

5) Enterprise knowledge intelligence: Bringing together internal and external knowledge, empowering employees, and providing expertise covering enterprise management and industry.

At the same time, the Sky Agent platform ushered in a major upgrade to version 2.0. Currently, it can provide richer templates and tools, bring a more integrated and unified SaaS experience, and make the models/algorithms/data more secure. The Sky Agent development platform also supports the integration of third-party tools or agents through protocols such as MCP and A2A, and is currently connected to ByteButton Space, Dify, etc.

In addition, Kingdee and Volcano Engine have closely cooperated in implementing AI. Among them, at the bottom level of computing power, Volcano Engine provides computing power support for Kingdee's AI business; in the middle layer, Kingdee supports the use of self-developed models and various tripartite open source models, including Doubao. The two parties will cooperate deeply at the platform layer and tool layer, including multi-modal model combinations, MCP fields and scenario optimization; at the upper level, based on Kingdee's accumulation in enterprise management scenarios such as finance, HR, supply chain, and office collaboration, combined with Volcano Engine's support at the model, tool, and computing power levels, it will better serve enterprise customer AI implementation.

Kunlun World Wide: Tiangong Super Intelligent System Brings Efficient Office Experience

On May 22, Kunlun World Wide simultaneously released Skywork Super Agents (Skywork Super Agents) to the global market. Skywork consists of 5 expert agents and 1 general agent. The former focuses on generating word/PPT/Excel/podcasts and web pages, while the latter is connected to dozens of MCPs to process and generate images, posters, music, MVs, promotional videos, audiobooks, picture books, and other multimedia content. At the same time, Kunlun World Wide has been thoroughly refined in terms of the specific implementation of the scenario. 1) When receiving user requirements, Skywork will require users to select goals and requirements through “clarification cards” and generate a “to-do list” for users to confirm twice to avoid inefficiency caused by not being able to clearly understand user intentions; 2) Skywork provides a “traceability” function, and all generated content can provide a clear and traceable source to avoid AI illusion; 3) Skywork also supports users to upload their own files and build a knowledge base to create a source of personal information and avoid interaction The generation quality is poor due to sources with uneven network quality. Based on these capabilities, Skywork scored 82.42 points in the GAIA evaluation set, surpassing OpenAI Deep Research and Manus, ranking first (data from May 10, 2025); it scored 94.5 points on the agent evaluation task SimpleQA launched by OpenAI, and achieved a new SOTA. Currently, Skywork supports direct registration and use by users at home and abroad, bringing users an efficient office experience.

Risk Alerts

(1) Downstream macroeconomic risk: The downstream of the computer industry involves thousands of industries. Under downward macroeconomic pressure, industry IT expenditure falling short of expectations will directly affect computer industry demand; (2) Bad accounts receivable risk: most computer companies' business is mainly project-based and can receive repayments after approval. Longer downstream customer payment cycles may increase bad accounts receivable and may further lead to asset impairment losses; (3) Increased competition in the computer industry: demand in the computer industry is more determined, but supply-side competition intensifies or causes changes in the industry pattern; (4) International The impact of environmental changes: International trade frictions are intensifying, and the US continues to put pressure on Chinese technology, which may have an impact on companies with a high share of overseas revenue.