Adopt
Agent Skills (SKILL.md) package procedural knowledge - instructions, scripts, resources - into a folder an agent loads on demand. Released by Anthropic as an open spec in December 2025, it became the fastest cross-vendor standard in AI tooling: within months 30+ tools read the same format, including Claude Code, OpenAI Codex, Cursor, VS Code, Gemini CLI and Goose. Skills complement MCP: a skill encodes how to do something, MCP provides access to tools and data.
Using skills well:
- Treat skills as code: versioned, reviewed, in a shared team repo - they are part of your harness (see Harness Engineering) and a natural self-improvement loop (agents can write and refine their own skills).
- Keep them small and focused: progressive disclosure is the point - the agent loads a skill only when needed, keeping context lean.
- Finding skills: anthropics/skills (official, curated), skills.sh (large community marketplace), plus the agentskills.io ecosystem. Curate before adopting.
- Security is the catch: skills are a new supply chain - third-party instructions and scripts running with your agent's permissions. When Snyk scanned ~4,000 published skills (Feb 2026), 36.8% had at least one security flaw, 13.4% a critical one. Scanners exist (Snyk Agent Scan, Cisco Skill Scanner, OSS skill-audit) - use them pre-install, but research shows they can be bypassed, so a green checkmark does not replace reviewing what you install. See Prompt Injection Awareness.
Recommendation
Adopt the format - it is the standard way to give agents reusable procedures across all major harnesses. Build an internal, reviewed skill repository instead of installing from marketplaces ad hoc, and scan third-party skills before use.