Research: Modular Agents with Acquirable Skills
A new paper explores a shift from monolithic LLMs to modular agents equipped with specific, acquirable skills. The research addresses the architectural, acquisition, and security challenges of building ecosystems of more specialized agents. This modular approach is intended to create more capable and secure systems compared to single, general-purpose models.
- The shift to modular agents is formalized in the paper through portable skill definitions (`SKILL.md`), progressive context loading, and integration with the Model Context Protocol (MCP), an open standard for connecting agents to external tools. This creates a stack where skills provide procedural knowledge ("what to do") and MCP handles data connectivity ("how to connect"). - A significant security challenge has emerged; recent analyses reveal that 26.1% of community-contributed agent skills contain vulnerabilities, prompting the paper's proposal of a four-tier, gate-based "Skill Trust and Lifecycle Governance Framework". - In China, major tech companies like Tencent, Alibaba, and ByteDance are embedding agentic AI directly into their "super-app" ecosystems, focusing on task completion and e-commerce rather than just conversational interaction. Tencent's Agent Runtime, for example, handles billions of tool calls daily within WeChat for tasks like customer service and accounting. - Open-source multi-agent orchestration frameworks are central to building these systems, with Microsoft's AutoGen noted for its flexible, chat-centric model and CrewAI for its simpler, role-playing approach to agent collaboration. Other prominent frameworks include LangGraph for stateful workflows and Google's Agent Development Kit (ADK). - Key architectural patterns for coordinating multiple agents include hierarchical (manager-worker), peer-to-peer, and market-based systems where agents might bid for tasks. Google has outlined eight essential design patterns, including sequential pipelines, parallel fan-out/gather, and hierarchical task decomposition, to provide structure for developers. - A primary technical hurdle in multi-agent systems is the non-linear increase in coordination overhead; every inter-agent handoff adds latency, and context-sharing costs can scale exponentially, potentially making a multi-agent system 3-5 times more complex to develop than a single-agent equivalent. - For consumer-facing products, user experience (UX) design is shifting to accommodate agentic systems, requiring new patterns like transparently displaying the AI's decision-making process, providing clear safety controls, and managing conversational history to maintain context. The design focus is moving from direct user control to outcome-oriented collaboration between the human and the agent. - The Chinese AI agent market reached a revenue of USD 577.0 billion in 2025 and is projected to grow to USD 14,796.0 billion by 2033. This growth is driven by companies like Baidu (Wenxin ecosystem), Ant Group (Lingji platform), and startups like Manus and Zhipu AI, which are developing general-purpose and vertical-specific agents.
