The agent stack, top to bottom
Tools, skills, agents — the three-layer mental model that explains every multi-agent system I've reviewed.
I’ve reviewed enough enterprise agent builds to know they fail in the same places. The failures aren’t about model quality. They’re about layer confusion.
Three layers, in order
Tools
Raw capabilities. The model, the APIs, the MCP servers, the vector store. This is the layer where vendors sell you things. It’s also the layer that’s been commoditizing fastest. Most companies overspend here because it’s the part you can shop for.
Skills
Encoded workflows. Reusable prompts. Tested chains. Named playbooks. This is where the company’s IP starts to live — in the way it stitches the tools together for its specific work.
Skills are where most enterprise builds get stuck. Engineers want to ship “the agent.” But without a skill layer, every agent is a snowflake. Every prompt is a copy-paste. Every fix has to be re-applied across implementations.
Agents
Persistent, role-defined systems that operate in the business. They use skills. They invoke tools. They have audit logs. They have a manager — sometimes human, sometimes another agent.
This is what most companies want when they say “AI strategy.” But you can’t build an agent layer that works without a skill layer underneath, and you can’t build skills without good tools.
What this means in practice
Map the stack you actually have. If you’ve been buying tools and trying to assemble agents directly, you have a middle-layer problem. If you’ve got Skills but no agents, you have an orchestration problem. If you’ve got agents but no Skills, you’ve built a snowflake that no one will be able to maintain when you leave.
The middle layer is where the work is. The middle layer is also where most of the value is.
(Essay-length take — more soon.)