Anthropic: The industry got AI agents wrong — a different approach will lead to a breakthrough

Speaking at the AI Engineering Code Summit, researchers Barry Zhang and Mahesha Murag said the key to effective automation is not increasing the number of agents but using a single, universal agent that can access a library of skills.

Skills are structured files — modular work instructions that contain everything an agent needs to perform tasks accurately and consistently.

According to Zhang, today’s agents «still lack expertise» and often miss important context when handling real-world cases. Skills help bridge this gap by providing agents with domain awareness and access to formalized workflows.

Murag noted that within the first five weeks after Anthropic’s tool release, users created thousands of such skills. Moreover, employees at major companies — in accounting, legal, and HR departments — are already using skills as internal operational instructions. Fortune 100 companies have begun implementing them to transfer corporate standards and best practices to agents, he added.

The skills concept emerges against the backdrop of a global race to develop AI agents. Tech executives regularly describe agents as the future of office work. OpenAI CEO Sam Altman has stated that agents are already performing entry-level employee duties and will eventually be able to seek new knowledge and solve complex business problems. Microsoft suggests that agent systems might even change corporate hierarchies.

But skepticism is growing. A16z co-founder Guido Appenzeller previously noted that many startups misuse marketing tactics — for example, adding a chat interface to a model and calling it an «agent» to justify higher prices.