Research Engineer at Anthropic working on Claude’s agent architecture and the Skills system. Co-author of influential articles on building effective agents and the Skills framework that enables agents to dynamically load procedural knowledge.
Champion of Unified Agent Architecture
Barry Zhang advocates for a paradigm shift: rather than building multiple specialized agents, organizations should build one universal agent powered by a library of Skills. This approach simplifies architecture while managing complexity through progressively-disclosed, shareable knowledge packages.
Current Work
As a Research Engineer at Anthropic, Barry focuses on:
- Claude Agent Architecture - Designing and implementing Claude’s core agent systems
- Skills System Development - Building the Skills framework that packages procedural knowledge agents can dynamically load
- Production-Grade Systems - Scaling agent systems for enterprise deployment
- Knowledge Management - Creating institutional knowledge repositories through Skills
Barry previously worked on Meta’s knowledge graph team, bringing deep expertise in large-scale data systems to his work on agent development.
Key Publications
Barry has co-authored influential articles that shape how the industry thinks about agents:
Building Effective Agents (December 2024, with Erik Schluntz) - Challenges the industry’s over-engineering approach, advocating for simple composable patterns over complex frameworks. Introduces the distinction between workflows and agents, emphasizing that “success in the LLM space isn’t about building the most sophisticated system. It’s about building the right system for your needs.”
Equipping agents for the real world with Agent Skills (October 2025, with Keith Lazuka and Mahesh Murag) - Introduces the Skills framework: organized folders of instructions, scripts, and resources that transform general-purpose agents into specialized agents through dynamic loading.
Philosophy on Agent Development
Barry’s approach emphasizes:
Universal agents over specialized ones - “The agent underneath is more universal than we thought.” One well-architected agent with comprehensive skills beats multiple specialized agents.
Progressive disclosure - Skills reveal only surface-level information initially, then provide full capability when needed, preventing context window explosion.
Institutional knowledge - “Skills collect institutional knowledge,” making agent memory tangible and shareable across organizations.
Simplicity wins - “Code is all you need” in the skills framework. Avoid complex multi-agent architectures when a unified agent with rich skills suffices.
Conference Appearance
Event: AI Engineering Code Summit 2025 Date: November 21, 2025 Time: 9:11 AM - 9:30 AM Session: Don’t Build Agents, Build Skills Instead Co-Presenter: Mahesh Murag
Barry and Mahesh presented on a fundamental shift in agent development: “We think we’ve converged on the architecture to build agents.” The session covered how thousands of Skills have been deployed in the last five weeks, demonstrating rapid scaling across domains from document processing to scientific research to enterprise workflows.
Key Insights
The Skills ecosystem has revealed important trends:
- Non-developer contributors building high-value professional skills
- Fortune 100 adoption with organization-wide skill libraries serving thousands of engineers
- Production-grade complexity as Skills mature into enterprise-ready systems
- MCP complementarity where Skills work alongside Model Context Protocol tools
Barry emphasizes that Skills make the concept of memory more tangible, providing structure needed for production AI systems while enabling shared knowledge infrastructure across organizations.