Exam-style
A team's internal assistant is built as an agent. It raises tickets, searches a document store, and looks up leave balances in an HR API, and each of those three connections was hand-written: a bespoke tool definition, a bespoke adapter, a bespoke bit of glue. A fourth system is now on the list, and a second team wants the same three. Someone suggests Model Context Protocol [MCP]. What does adopting it give them?
Reveal the answer
B. An open protocol for connecting agents to external tools and data sources through one common interface, so a system exposed once can be called by any MCP-compatible agent
MCP is an open protocol that standardises how an agent reaches things outside itself. A system publishes its capabilities once, as an MCP server, describing what each tool does and expects; any MCP-aware agent can then discover and call those tools through the same interface. The fourth system gets exposed once rather than wired into one agent, and the second team points its agent at servers that already exist. What it does not change is tool selection: a badly described tool is still selected wrongly, however it is connected. It is not a prompt format. It is not a memory store; what an agent keeps within and between conversations is memory management, with its own retention rules. It is not agent-to-agent plumbing; delegation, handoff and a shared workspace are multi-agent communication patterns, and MCP runs the other way. Nor is it proprietary, though AWS does implement it: Amazon Bedrock AgentCore Gateway exposes APIs a company already has as MCP tools, so an existing HR API becomes callable without a hand-written adapter.
Q. An agent reaches a ticketing system, a document store and an HR API, and all three connections were hand-written. A fourth system is coming and a second team wants the same three. What does adopting Model Context Protocol [MCP] give them?
A. One common interface between agents and external systems. A system is exposed once as an MCP server, and any MCP-compatible agent can discover and call its tools without a bespoke adapter per pairing.
Why? Sort the confusions by direction and by layer. MCP runs outward, from one agent to the systems it acts on, so it is not the agent-to-agent plumbing; delegation, handoff and a shared workspace are multi-agent communication patterns. It sits at the connection layer, not the prompt layer and not the memory layer, so it is neither a way of writing prompts nor a place to keep conversation history. And it is an open standard rather than an AWS product, which is why the AWS piece is a service that implements it: Amazon Bedrock AgentCore Gateway exposes existing APIs as MCP tools. Describing each tool well enough that the right one gets selected is still the work, because those descriptions are what the model selects from.