Anthropic Engineers Say Stop Building Agents, Build Skills Instead

Barry Zhang and Mahesh Murag, the engineers who built Claude Skills at Anthropic, published a talk and engineering post that's gotten 14K+ likes and is reshaping how I think about agent development. The core argument: most agent approaches fail because they lack domain expertise. The fix isn't better models. It's better skills.

Skills are organized folders of instructions, scripts, and resources that agents discover and load dynamically. The key design principle is progressive disclosure. Claude loads information only as needed rather than consuming entire skill definitions into context. Anthropic says they now run hundreds of skills in production.

The same week, Anthropic released Agent Skills as an open standard with launch partners including Canva, Notion, Figma, and Atlassian shipping prebuilt skills. Enterprise admins on Team and Enterprise plans get centralized skill management. This creates a formal plugin ecosystem for Claude's agentic capabilities.

The ecosystem response has been immediate. VoltAgent's awesome-agent-skills repo now has 1,000+ hand-curated skills from Anthropic, Google Labs, Vercel, Stripe, Cloudflare, and 40+ other teams. HuggingFace published SKILL.md files that teach coding agents to write production CUDA kernels. Both Claude Code and Codex successfully produced working kernels with correct PyTorch bindings using these skills.

This lands for me because I've been writing skills for my own Claude Code setup for months. The difference between a naked agent and one loaded with the right procedural context is enormous. I've watched the same model go from mediocre output to production-ready code just by giving it a well-structured SKILL.md that encodes the decisions I'd make myself.

The reframe is important. "How do I build an agent" is the wrong question. "How do I encode procedural knowledge so any agent can use it" is the right one. Skills are portable across models and tools. They work with Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot, and Windsurf. Your investment in writing a good skill survives the next model upgrade.

For builders: start writing skills today. Pick the workflow you repeat most often. Document the decisions, the gotchas, the specific commands and file patterns. Put it in a SKILL.md. Test it with your agent of choice. You'll get more mileage from one well-written skill than from switching models.