AI Agents for Beginners - 11. Using Agentic Protocols (MCP, A2A and NLWeb)

AI Agents for Beginners - 11. Using Agentic Protocols (MCP, A2A and NLWeb)

Source: microsoft/ai-agents-for-beginners - 11-agentic-protocols

Model Context Protocol

MCP Core Components

MCP Client-Server Architecture
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart LR
    subgraph Host["Host Application"]
        direction TB
        LLM["LLM"]
        Client["MCP Client"]
        LLM <--> Client
    end
    subgraph Server["MCP Server"]
        direction TB
        Tools["Tools\n(Actions)"]
        Resources["Resources\n(Data)"]
        Prompts["Prompts\n(Templates)"]
    end
    Client -->|"discover capabilities"| Server
    Server -->|"tool list + schemas"| Client
    Client -->|"call_tool(name, params)"| Server
    Server -->|"tool result"| Client

Benefits of MCP

MCP Example

1. Connection: The AI assistant (MCP client) connects to an MCP server provided by an airline.
2. Tool Discovery: The client asks the airline's MCP server, "What tools do you have?" The server responds with a list of tools such as search_flights and book_flights.
3. Tool Invocation: When requested to "Search for flights from Portland to Honolulu," the AI assistant uses the LLM to determine that it needs to call the search_flights tool and passes the relevant parameters (origin, destination) to the MCP server.
4. Execution and Response: Acting as a wrapper, the MCP server calls the airline's internal booking API. It receives the flight information (JSON data) and returns it to the AI assistant.
5. Further Interaction: The AI assistant presents flight options. Once a flight is selected, the assistant calls the book_flight tool on the same MCP server to complete the booking.

MCP Example: Flight Booking
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart LR
    User(["User"])
    Agent["AI\nAgent"]
    MCP["MCP\nServer"]
    API["Booking\nAPI"]

    User -->|"③ Request search_flights"| Agent
    Agent -->|"① Connection"| MCP
    Agent <-->|"② Tool Discovery\nsearch_flights, book_flights"| MCP
    MCP -->|"④ Execution\nCall internal API"| API
    API -->|"④ Response\nFlight JSON"| MCP
    MCP -->|"④ Returns\nFlight Options"| Agent
    Agent -->|"④ Shows\nFlight Options"| User
    User -->|"⑤ Selects Flight\n→ Call book_flight"| Agent

Code: MCP-Style Tool Discovery

setup.py
python

mcp_tools.py
python

mcp_agent.py
python

MCP Agent: Tool Discovery Flow
Mermaid
flowchart LR
    User["User Query\n'Tokyo 5 nights, culture & food'"]
    --> Agent["AccommodationAgent\n(MCP Client)"]
    Agent -->|"discover tools"| MCP["MCP Server\n(Simulated)"]
    MCP -->|"tool list + schemas"| Agent
    Agent --> T1["search_accommodations\n(Tokyo, April, 2 guests)"]
    Agent --> T2["get_local_experiences\n(Tokyo, culture + food)"]
    T1 & T2 --> LLM["LLM\nSynthesize & compare results"]
    LLM --> Response["Personalized accommodation\n& experience recommendations"]
    Response --> User

Agent-to-Agent Protocol (A2A)

A2A Core Components

A2A Core Components
Mermaid
flowchart TD
    Orch["Orchestrator Agent\n(Travel Agent)"]
    Orch -->|"Look up Agent Card"| Registry["Agent Card Registry\nName · Skills · Endpoints"]
    Registry -->|"Discovered agents list"| Orch
    Orch -->|"Pass context"| AE["Agent Executor\nExecute remote agent"]
    AE -->|"Task complete"| Art["Artifact\nOutput · Explanations · Text"]
    Art --> Orch
    AE <-->|"Progress updates"| EQ["Event Queue\nMaintain connection · Message delivery"]

Benefits of A2A

A2A Example

A2A Example: Full Trip Booking
Mermaid
flowchart TD
    User["User\n'Book round-trip flight + hotel + car rental for Honolulu'"]
    --> TA["Travel Agent\n(A2A Orchestrator)\nAnalyze task with LLM"]
    TA -->|"A2A Protocol"| AA["Airline Agent\n(Different company)\nSearch & book flights"]
    TA -->|"A2A Protocol"| HA["Hotel Agent\n(Different company)\nBook hotel"]
    TA -->|"A2A Protocol"| CA["Car Rental Agent\n(Different company)\nBook rental car"]
    AA & HA & CA -->|"Return Artifact"| TA
    TA -->|"Aggregate results"| Response["Consolidated Travel Confirmation\nIntegrated flight, hotel, and car rental"]
    Response --> User

Code: A2A Multi-Agent Workflow

a2a_agents.py
python

a2a_workflow.py
python

A2A Workflow: 3-Agent Pipeline
Mermaid
flowchart LR
    User["User\n'1-week trip to Tokyo\nInterested in food, temples, technology'"]
    --> CE["CurrencyExchangeAgent\nExchange rates & currency guide\nBest timing & tips"]
    CE -->|"Pass context"| AP["ActivityPlannerAgent\nRecommend attractions, stays, dining\nTailored to traveler interests"]
    AP -->|"Pass context"| TM["TravelManagerAgent\nSynthesize full itinerary\nGenerate structured travel brief"]
    TM --> Response["Final Travel Plan\nExchange rates + Activities + Comprehensive brief"]
    Response --> User

Natural Language Web (NLWeb)

Components of NLWeb

NLWeb Architecture
Mermaid
flowchart TD
    Content["Website Content\n(Schema.org / RSS Feeds)"]
    --> Ingest["NLWeb Application\nData ingestion & processing"]
    --> Embed["Embedding Model\nConvert content → vectors"]
    --> VDB["Vector Database\n(Qdrant / Azure AI Search, etc.)"]

    Query["Natural Language Query\n(User or AI Agent)"]
    --> NLP["LLM\nQuery understanding & interpretation"]
    --> Search["Vector Search\nRetrieve relevant embeddings"]
    Search --> VDB
    VDB -->|"Return similar results"| NLP
    NLP --> Response["Natural Language Response\n(Grounded in real DB data)"]

    NLWeb_MCP["NLWeb as MCP Server\n'ask' method"] -->|"Called by AI Agent"| Search

NLWeb by Example

1. Data Ingestion: The travel website's existing product catalog (flight listings, hotel descriptions, tour packages, etc.) is formatted into Schema.org structures or ingested via RSS feeds. An NLWeb tool converts this structured data into embeddings and stores them in a local or remote vector database.
2. Natural Language Query: Instead of navigating through menus, a user visiting the website types into a chat interface: "Find a family-friendly hotel in Honolulu with a pool for next week."
3. NLWeb Processing: The NLWeb application receives the query. It sends the query to an LLM to comprehend its intent while simultaneously querying the vector database for relevant hotel listings.
4. Accurate Results: The LLM interprets the database search results, selects optimal matches based on "family-friendly", "pool", and "Honolulu", and generates a natural language response. Because it references actual hotels from the website's catalog, hallucinations are avoided.
5. AI Agent Interaction: Since NLWeb also functions as an MCP server, external AI travel agents can connect directly to the website's NLWeb instance. It queries directly via the ask MCP method: ask("Are there any hotel-recommended vegan restaurants in Honolulu?") → NLWeb processes this and returns a structured JSON response.

NLWeb Example: Travel Booking Site
Mermaid
flowchart TD
    S1["1. Data Ingestion\nFormat flights, hotels, & tour packages\nusing Schema.org/RSS\nGenerate embeddings → Store in Vector DB"]
    --> S2["2. Natural Language Query\nUser: 'Family-friendly Hawaii hotel with pool'\nChat input without menu navigation"]
    --> S3["3. NLWeb Processing\nLLM: Understand query\nVector DB: Search relevant hotels"]
    --> S4["4. Accurate Results\nLLM interprets results & selects best matches\nResponse based on real catalog\n(No hallucinations)"]
    --> S5["5. AI Agent Interaction\nConnect external AI Agent via MCP 'ask' method\nask('Hawaii vegan restaurants?')\n→ Structured JSON response"]

Summary

Lesson 11 Summary: Three Agentic Protocols
Mermaid
flowchart LR
    Root["Agentic\nProtocols"]
    Root --> MCP["MCP\nModel Context Protocol"]
    Root --> A2A["A2A\nAgent-to-Agent"]
    Root --> NLW["NLWeb\nNatural Language Web"]

    MCP --> M1["Standardized Tool Access"]
    MCP --> M2["Dynamic Discovery"]
    MCP --> M3["Tools · Resources · Prompts"]

    A2A --> A1["Agent-to-Agent Collaboration"]
    A2A --> A2["Agent Card · Artifact"]
    A2A --> A3["Cross-framework Interop"]

    NLW --> N1["Websites + Natural Language Interface"]
    NLW --> N2["Vector DB-based Search"]
    NLW --> N3["Also Functions as an MCP Server"]

• MCP serves as a "universal adapter" that grants AI Agents standardized access to external tools and data, allowing dynamic discovery of new capabilities without altering agent code.
• A2A enables agents from different organizations and frameworks to discover one another via Agent Cards and collaborate to complete complex tasks.
• NLWeb adds a natural language interface to websites and functions as an MCP server, empowering external AI Agents to query web content directly via the ask method.
• The three protocols complement one another: MCP handles tool integration, A2A drives agent-to-agent collaboration, and NLWeb bridges web ecosystems with AI Agents.