Over the past year, the technology sector has engaged in an intense and highly competitive race to design and deploy increasingly sophisticated artificial intelligence agents. These software entities are envisioned as autonomous digital assistants capable of performing a wide range of professional tasks once reserved for human workers. Yet, despite this widespread enthusiasm, researchers at Anthropic contend that the industry’s obsession with proliferating more agents may be misguided. They argue that a more elegant and ultimately more impactful concept—centered on what they describe as ‘agent skills’—could significantly enhance the real-world performance and reliability of AI systems.
During the AI Engineering Code Summit held last month, Anthropic researchers Barry Zhang and Mahesh Murag elaborated on this crucial shift in perspective. They asserted that the true milestone in optimizing agent workflows will not emerge from producing an endless variety of specialized agents, but rather from cultivating a shared repertoire of refined skills that any agent can draw upon. According to Zhang, who spoke in a discussion later published online, the company’s research suggests that while agents may appear diverse across different domains, their core architecture is far more universal than previously assumed. This discovery implies that, instead of inventing a unique agent for every conceivable scenario, it is more efficient to employ a single, adaptable agent model equipped with an extensive library of modular skills.
Zhang described these skills as organized and comprehensively structured collections of files. Each set functions much like a self-contained package that consolidates procedural knowledge, best practices, and step-by-step operational logic. In simpler terms, a skill is a digital toolkit—a folder containing everything an agent needs to execute its assigned task with accuracy, consistency, and speed. Through these organized skill sets, agents can replicate human-style expertise and procedural understanding without needing to reinvent their internal functions every time they encounter a new task.
Despite their rapidly advancing capabilities, Zhang emphasized that most AI agents currently fall short in terms of specialized expertise. They frequently fail to capture the subtle contextual nuances that are fundamental to solving real-world business problems. This limitation often results in outputs that, while technically correct, lack the depth of understanding a domain expert would provide. Skills, therefore, serve as the mechanism that bridges this crucial gap, supplying agents with embedded domain intelligence and reusable operational workflows that simulate human professional knowledge.
Murag provided further insight into how this concept is already being adopted in practice. He noted that individuals across non-technical departments—such as accounting, human resources, legal, and recruiting—are actively developing and sharing their own sets of skills. In the short period of just five weeks following the feature’s launch, thousands of these reusable skills were created. Major corporations have begun integrating them into internal systems, treating each skill not merely as a technical asset but as a form of institutional knowledge—a codified ‘playbook’ that reflects how the company prefers to operate its AI tools. According to Murag, Fortune 100 companies are already leveraging these capabilities to teach agents about their organizational standards, internal policies, and best operational practices.
The discussion surrounding agent skills takes place within a broader context often referred to as ‘the rise of AI agents.’ Across the technology ecosystem, industry leaders have heralded these agents as potential catalysts for a fundamental transformation in modern office work. OpenAI CEO Sam Altman, for instance, remarked at the Snowflake Summit 2025 that AI agents are beginning to handle a range of tasks traditionally assigned to entry-level employees. He illustrated how many professionals now find themselves managing teams of agents—assigning projects, evaluating their performance, and integrating their results—much as they would when supervising junior colleagues. Altman also predicted a near future in which these systems will evolve from routine task-handling to more advanced forms of reasoning, where they might contribute to the discovery of new knowledge or the formulation of creative solutions to complex business challenges.
Asha Sharma, who leads Microsoft’s AI platform product strategy, echoed this transformative vision during an episode of ‘Lenny’s Podcast’ released in August. She proposed that the growing presence of effective AI agents could eventually reshape corporate hierarchies, reducing the necessity for multiple managerial layers. In her view, organizations could begin to operate with flatter, more agile structures as AI agents take on many of the coordination and communication tasks that currently require human intermediaries.
However, not everyone in the technology sector shares this optimistic outlook. Some veteran investors and engineers have voiced concerns that the current excitement around AI agents has spiraled into hype. Guido Appenzeller, a partner at the venture capital firm Andreessen Horowitz (a16z), observed in a company podcast that a number of startups are using the term ‘agent’ primarily as a marketing label. According to him, some firms are rebranding simple chat-based interfaces built on large language models as fully autonomous agents, enabling them to justify higher pricing and attract investor interest under a trendy buzzword. He cautioned that while the agent concept holds promise, its commercial overuse risks diluting its true value by turning it into a branding exercise rather than a technological breakthrough.
Taken together, these perspectives reveal a growing tension between quantity and quality in the evolution of AI. The industry’s initial instinct—to generate more agents in pursuit of greater capability—is being challenged by Anthropic’s rationale that true advancement lies in teaching each agent how to learn, adapt, and perform more intelligently through an expandable repository of skills. If this vision prevails, the next major leap in artificial intelligence may not come from building endless new entities but from equipping existing ones with the cognitive tools they need to work like seasoned professionals rather than mere digital interns.
Sourse: https://www.businessinsider.com/anthropic-researchers-ai-agent-skills-barry-zhang-mahesh-murag-2025-12
