Agents or Skills, industry is evolving?

I build fewer and fewer custom agents. Today’s default agents are already very capable. What they miss is my context.

That context now goes into skills.

Three things I see in my own work:

→ Skills complement agents. They don’t replace them. Agents still matter where the focus is the goal. Skills are about how to do the work.

→ Learn from public skills, but build your own. I don’t trust marketplace skills yet, unless they come from a reputed company with robust certification. Even big companies’ CI/CD pipelines have been compromised, and a skill runs with the agent’s permissions. So I learn from them and build my own. Like a person learning a new skill.

→ Specs make vibe coding controlled. I was intrigued when GitHub Spec Kit moved from agents to skills. Spec, plan and tasks are written down, so the knowledge stays with the project, not in a chat window.

My advice for companies starting now:

Give your agent a job success profile. ▸ Skills ▸ Behaviours ▸ Key measures ▸ Stakeholders ▸ Responsibilities

Keep these outside the agent, so they can live and grow on their own.

Agents define the goal. Skills define how.

(MCP is a different topic. It is about connecting agents to tools and data, so I keep it for another post.)

Are you also building more skills and fewer agents?

#AgenticAI #AIAgents #AgentSkills #SpecDrivenDevelopment #SoftwareEngineering