Torii Report: AI Is Expanding Corporate 'Shadow IT'

- The average enterprise now contends with over 830 applications, a figure that can escalate to 2,191 for large enterprises, with the average employee using 40 different applications. Only 15.5% of these applications are formally sanctioned by IT departments, leaving the rest in states like 'in review,' 'blocked,' or completely unknown. - To manage the complexity of coordinating multiple AI agents, which is a core challenge in moving from single-task bots to enterprise-grade systems, open-source frameworks like Microsoft's AutoGen and CrewAI are gaining traction. AutoGen focuses on flexible, chat-centric orchestration, while CrewAI emphasizes a role-based approach to simplify the definition of agents and tasks. - Architectural patterns for multi-agent systems are moving beyond simple sequential workflows to more complex designs like the "coordinator pattern," where a central agent decomposes tasks and dispatches them to specialized agents. Other patterns include parallel execution for simultaneous independent analysis and hierarchical structures for breaking down complex problems into manageable sub-tasks. - Production deployments of multi-agent systems reveal significant reliability challenges not often found in single-agent setups. Core failure points include state synchronization errors, where agents act on outdated information, and communication protocol breakdowns, where message ordering violations lead to cascading errors. A key metric is "coordination overhead," which can scale exponentially if not architected correctly. - A 2026 research paper on "Dynamic Planning and Tool Use in Next-Gen AI Agents" highlights a critical shift from static prompting to agentic frameworks where LLMs autonomously decompose tasks, select tools (like APIs or databases), and adapt reasoning in real-time. This approach demonstrates significant performance gains in reliability and task-completion rates over traditional methods. Additional 2025 research focuses on standardizing agent-to-agent communication protocols and developing frameworks for evaluating multi-agent collaboration. - In China, the AI agent landscape is rapidly advancing with major players like Baidu (ERNIE Bot), Tencent (Hunyuan), and iFlytek (SparkDesk) offering agent development platforms. At the same time, startups like Genspark and Manus are gaining global attention, with Genspark reportedly reaching $10 million in annual recurring revenue in just nine days. - The regulatory environment in Beijing is solidifying around a "local-first" principle, which shapes AI model architecture and data strategies for companies operating in China. New rules that took effect in late 2025 and saw intensified inspections in early 2026 mandate visible and hidden labels on all AI-generated content to increase transparency and control.