Build Skills, Not Agents

For two years, every team building with LLMs has been reinventing the same wheel: the agent loop. Custom harnesses, custom orchestrators, custom retry logic, custom tool routers. Meanwhile, Claude Code, the Claude Agent SDK, OpenAI's Agents SDK, and a handful of mature open frameworks have quietly converged on the same pattern. The agent harness is, for most practical purposes, solved.

So if the harness is solved, what's left to build? Enter Agent Skills.
What differentiates my agent from yours is not the loop — it is the skills and tools we equip it with. A generic agent plus the right skills becomes a legal analyst. The same agent plus a different set of skills becomes a deployment engineer, a financial controller, or a research assistant. Skills are how you go from a tool-calling chatbot to an agent that actually does the job.

In this talk, we will use the Claude Agent SDK as our harness — solved, off-the-shelf, no reinvention required — and focus all our energy on the layer that actually matters. You will get a clear mental model of the skill + tool stack, live demos of building skills two ways (with the skill creator pattern, and by capturing real workflows as reusable capabilities), and a practical playbook for composing skills with MCP servers and CLI tools to build the "everything agent" your team actually needs.

Stop building harnesses. Start building skills.

NDC { Oslo }

Build Skills, Not Agents

Aleksander Stensby