Barry Zhang

Barry Zhang is a member of Anthropic’s technical staff and a co-creator of Agent Skills, a framework that gives general-purpose AI agents specialized expertise through reusable files. He focuses on making agents reliable enough for consequential work without unnecessary complexity, cost, or autonomy.

From practical AI products to agent architecture

Zhang studied computer science and industrial engineering at Northwestern University and previously worked at Meta, where he was a technical lead on its generative-AI monetization team. While building AI products there in 2023, he adopted the emerging identity of an AI engineer, drawn to its emphasis on practical usefulness.

His earlier projects anticipated several directions in applied AI: BKMK uses retrieval-augmented generation to answer questions about saved bookmarks, while Auto-distill-GPT explores language-model knowledge distillation. Under his full name, Yijing Barry Zhang, he coauthored research on fine-tuned language models for text classification, including intent detection and slot filling.

At Anthropic, Zhang developed a practical philosophy of agent design before helping create Agent Skills with Mahesh Murag and colleagues. His priorities are consistent: choose autonomy carefully, give models usable tools, and preserve expertise in forms that people and agents can improve.

• Autonomy should justify its costs. In Building effective agents, coauthored with Erik Schluntz, Zhang distinguishes predefined workflows from agents that choose actions based on environmental feedback. Predictable tasks usually warrant explicit workflows; agents make more sense when valuable work requires judgment and results can be checked. Software development fits that description because unit tests, continuous integration, and human review make outputs verifiable. His own coding workflow starts with a design document and verified tests before delegating implementation.

• Agent-computer interfaces determine reliability. Zhang models an agent as a system using tools in a loop, shaped by its environment, instructions, and available actions. A computer-use agent given a screenshot without screen dimensions or clear tool descriptions can fail for reasons invisible to its developer. He recommends examining complete execution trajectories and improving the model’s context before adding elaborate orchestration.

• Agent Skills package portable procedural knowledge. Skills organize instructions, scripts, examples, and supporting resources into ordinary directories. One reusable presentation-styling script, for example, can replace a procedure an agent would otherwise rewrite repeatedly. Progressive context disclosure keeps these libraries manageable: the model initially sees concise metadata, loads detailed instructions when needed, and retrieves additional files selectively.

• Connectivity and expertise are separate layers. The Model Context Protocol connects agents to external systems; skills tell them how to perform useful work with that access. Zhang envisions shared, evolving libraries of organizational knowledge that agents can reuse across tasks and sessions. His public introduction to Skills describes this as a practical approximation of continuous learning. His account of multi-agent search adds a crucial constraint: impressive prototypes are comparatively easy; dependable production systems demand substantially more engineering.