FeedDiscussion
Kartik N V J K
AI Developer | Making AI reliable, trustworthy & accessible to everyone | Active community contributor
Really appreciate this breakdown. The three-layer model clears up a lot of the confusion around tools versus MCP versus skills. One thing I would add is that in practice, the boundaries get blurry at the edges. When a skill references a tool by name, it effectively becomes a routing layer, and MCP servers that return tool schemas dynamically are doing something closer to skill work than tool work. The real value of separating them becomes clear at scale: you can version your MCP servers independently from your tools, and your skills independently from both. That decoupling is what makes agent systems maintainable beyond a handful of functions. The tradeoff nobody talks about is latency. Every layer adds a round trip, and in a tight loop, those microseconds add up. Would love to see benchmarks comparing a single-tool MCP call against a native tool call in the same framework.
The one-sentence framing (tools are what, MCP is where from, skills are how) is the cleanest split I've seen and lines up with where my own bugs live. In practice the failure I keep hitting sits on the skills layer: two tools with near-identical descriptions, and the model picks the wrong one under load. Curious if you've found a routing/description-linting habit that catches those before they ship, because parameter-schema alone hasn't been enough for me.