CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
Week 5 you hand-wired tools into one agent. That works — once. Do it for five agents and twenty systems and you are maintaining a hundred pieces of custom glue. This week: the two open protocols that turn that multiplication into an addition — MCP for the tools and data underneath an agent, A2A for other agents beside it.
Move the sliders to size your organization, then flip between the two worlds. The red spaghetti is what every team builds first. The blue hub is what a standard buys you — the same connections, without the combinatorics.
Every red line is a bespoke integration somebody wrote, tested, and now owns.
connections to build and maintain — one per agent-system pair.
connections — M clients plus N servers, each written once.
Before USB-C, every device shipped its own charger and its own cable. The port did not make laptops better; it made everything else plug in. MCP is pitched the same way: one connector shape between an AI application and the tools and data it needs, so a connector built for one agent works for the next one. Anthropic used exactly this framing when it introduced MCP as an open standard in November 2024 (Anthropic, 2024 ↗).
A USB-C cable cannot lie to you. An MCP server can: it runs code, it returns text your model will read as instruction-shaped input, and it sits inside your trust boundary the moment you connect it. Plugging in a charger is a hardware decision; plugging in a server is a supply-chain decision. Hold onto that difference — it is the security thread that runs through page 01 and the case on page 03.
The single most useful thing you can memorize this week: MCP connects an agent to tools and data. A2A connects an agent to other agents. They are complementary, not competing — a single agent can speak both at once.
Your agent reaches down into a filesystem, a database, a ticketing API, a document store. The thing on the other end is not intelligent — it exposes capabilities and data, and your agent decides what to do with them. Open standard, introduced by Anthropic (2024); broadly adopted across the industry during 2025.
Your agent reaches sideways to another agent — possibly built by another team, another vendor, another company — and delegates a task rather than calling a function. The other side has its own goals, its own tools, and its own judgment. Announced by Google in April 2025 and contributed to the Linux Foundation in June 2025.
The line item that kills agent projects is rarely model spend — it is integration and the maintenance tail behind it. M+N is a budget argument before it is an architecture argument, and it is the one a CFO can check with arithmetic.
If your sixty connectors are written against one vendor's proprietary format, your renewal conversation has no exit. If they are written against an open protocol, switching hosts is a configuration change. Optionality is negotiating power.
A protocol boundary is also a control point: one place to log every call, allow-list every server, pin every version, and put the human gate. Ad-hoc glue has no such place — which is why nobody can tell you what it did last Tuesday.