What Agents Actually Need
The agent doesn't care about the protocol. It cares about getting things done.
Agents need
- A way to discover capabilities
- A way to invoke them reliably
- Structured output they can parse
- Error messages they can act on
- Documentation (a skill file, a man page, a
--help)
Agents don't need
- To speak JSON-RPC directly
- To manage MCP server lifecycle
- To hold open transport sessions (stdio, SSE)
- 50 tool schemas in context they'll never call
- Your team's bespoke MCP server implementation
Why this already works
- →Agents can write code — they already know how to call APIs
- →Agents can
exec() + attach stdio — they already know how to use CLIs
- →Models have 40+ years of training data on existing Windows CLIs
- →For new CLIs, a
skill.md file bridges the gap — the skill file is the protocol