We built apps for users. We perfected the GUI. Then we built services for other software, and we created REST, GraphQL and a pile of standards around them. Now the users of our software are AI agents. So: MCP?
APIs have become the backbone when designing our apps. MCP is a new standard of communication that lets agents use them efficiently, without guessing from the endpoints or reading the whole documentation every time an agent needs something. MCP solved a real problem and was quickly adopted across the industry.
Apps communication before MCP
Apps communication after MCP
MCP is not just a wrapper around an API. It is a proxy that helps the agent know how to ask for information and what to expect.
I am sure you have always found it hard to know what to keep in code, what to keep in documentation, and how some information ends up in neither. The accident that led a team member to make a change, a client’s specific need, a user story open to many readings…
I shall reconsider human knowledge by starting from the fact that we can know more than we can tell.
— Michael Polanyi, The Tacit Dimension (Chicago: University of Chicago Press, 1966), p. 4
Some knowledge is hard to write down. You get it only through a real conversation with a colleague or a friend. It is hard to put into an API or an MCP, but a skill can get close.
In practice, you soon see that agents sometimes don’t know when to ask, or why they need something. Like an explorer with a map who has never set foot in the jungle, the explorer does not know why a road is closed or why it needs another route. Only someone with experience can handle that. With AI agents, The documentation is the map: it shows which roads exist. The reason a road is closed usually lives in someone’s head. This is where the last piece of the architecture comes in: the Skill.
Skills have been adopted across the industry very fast and became a standard. They carry the information we cannot describe easily: the fine tuning of a need, it is like the experience of the explorer that guides us when using the map.
With my team, we work on an architecture that combines: a GUI for our users, built on the API (the engine), described by an MCP server, which talks to AI agents that have a skill, so they understand the app in depth.
The ideal harness architecture
This makes the skill part of the app, not an extra file, so updating it in every iteration is a must. In every version we update:
- Documentation describing the business need and where it lives in the app
- Code containing the feature (it must be self-explanatory for AI)
- The MCP and the skill
This is our small contribution to the dead internet prophecy, our agents make very few mistakes, consume less tokens and work faster.
Top comments (0)