Anthropic researchers Barry Zhang and Mahesh Murag said the AI industry should stop building large numbers of AI agents in favor of skill-first architectures [1].

This shift represents a fundamental change in how artificial intelligence is deployed. By moving away from complex agent-based systems, developers may be able to reduce system overhead and increase the reliability of AI outputs without the need for constant retraining.

The researchers presented their views last month at the AI Engineering Code Summit [1]. Their argument centers on the adoption of a simpler approach, such as the Model Context Protocol (MCP) [1, 2]. This framework allows AI systems to adapt and function based on specific skills rather than relying on the orchestration of multiple autonomous agents [2].

A skill-first approach is presented as more effective because it allows systems to evolve without the need to retrain large language models [1, 2]. This reduces the overall complexity of the AI stack, a significant hurdle for enterprises attempting to scale these technologies.

However, the industry remains divided on the utility of agents. While Zhang and Murag advocate for a move away from them, other perspectives suggest that the debate has shifted from whether to use agents to how to deploy them effectively [3]. This tension highlights a growing divide between the researchers focusing on architectural efficiency and the organizations focused on immediate implementation.

The viewpoint was published on May 15, 2026 [1, 3].

“The industry should stop building tons of AI agents and focus on simpler, skill‑first approaches.”

The proposal to move from 'agents' to 'skills' suggests a pivot toward modularity in AI. If the industry adopts the Model Context Protocol or similar frameworks, AI development may shift from creating autonomous entities to building a library of interoperable capabilities. This would lower the technical barrier for updating AI behavior, as developers could modify a specific skill rather than re-engineering an entire agentic workflow or retraining a massive model.