Skip to main content

Ticket Booking Example

The source files live in examples/ticket_booking.

This example demonstrates a practical multi-agent vacation planning workflow. A user asks for a relaxing Greek island trip, and a coordinator agent delegates to specialist agents for advice, weather validation, and hotel booking.

It highlights:

  • How a coordinator agent uses an LLM to plan and route work
  • How LLM-only and tool-only agents collaborate through agent_call
  • How agents discover each other through the Registry
  • How deterministic tools return structured booking and weather payloads
  • How the same scenario can run as an all-in-one quickstart or modular example

User Request​

Book me a relaxing vacation to Santorini for 5 nights in mid-July 2026.

From this single task, the system:

  1. Asks the Holiday Advisor for destination guidance.
  2. Checks weather through the Weather Agent.
  3. Books a hotel through the Hotel Agent.
  4. Returns a consolidated vacation summary to the user.

Agent Overview​

Coordinator Agent​

The coordinator is the primary entry point. It receives the user request, discovers available agents, and decides when to delegate.

  • Has an LLM
  • Uses agent_call with infer for advisory reasoning
  • Uses agent_call with tool_call for deterministic weather and hotel tools

Holiday Advisor Agent​

The advisor is an LLM-only specialist for travel recommendations.

  • Interprets vacation preferences
  • Evaluates destinations, dates, budget, and suitability
  • Returns concise structured advice to the coordinator

Weather Agent​

The weather agent is tool-only.

  • Owns a get_weather tool
  • Returns structured weather data for the selected destination and dates
  • Keeps external-data behavior deterministic and testable

Hotel Agent​

The hotel agent is tool-only.

  • Owns a book_hotel tool
  • Calculates nights and total price
  • Returns a structured booking confirmation

Registry​

The Registry stores each agent's AgentCard, including skills and capabilities, so the coordinator can discover peers dynamically.

Sequence Diagram​

Agent Classification​

AgentUses LLMHas ToolsPurpose
CoordinatorYesNoPlans, routes, and summarizes
Holiday AdvisorYesNoTravel reasoning and recommendations
Weather AgentNoYesWeather lookup
Hotel AgentNoYesHotel booking
RegistryNoNoDiscovery and registration

Files and Structure​

examples/ticket_booking/
├── quickstart.py             # All-in-one demo script
├── run.py                    # Modular demo entry point
├── coordinator_agent.py      # LLM-powered coordinator
├── holiday_advisor_agent.py  # LLM-only travel advisor
├── weather_agent.py          # Weather tool agent
├── hotel_booking_agent.py    # Hotel booking tool agent
├── .env.example              # Environment template
└── README.md                 # Example-specific README

Running the Example​

  1. Install Protolink with the required extras

    pip install "protolink[http,llms]"

  2. Choose an LLM provider

    === "Ollama"

    ollama pull gemma4:e4b
    ollama serve

    === "OpenAI"

    export OPENAI_API_KEY=sk-...

    === "Anthropic"

    export ANTHROPIC_API_KEY=sk-ant-...

  3. Run the all-in-one quickstart

    cd examples/ticket_booking
    python quickstart.py

  4. Or run the modular demo

    cd examples/ticket_booking

    # Ollama is the default provider
    python run.py

    # Use a hosted provider through environment variables
    LLM_PROVIDER=openai python run.py

    # Pass a custom request
    python run.py "Book a 5-night relaxing Santorini trip for two adults"
Registry startup

The checked-in quickstart.py and run.py scripts start the Registry and all agents for you. You do not need a separate registry terminal for this example.

Expected Result​

The exact prose depends on the selected LLM, but the final response should include:

  • Destination recommendation
  • Weather suitability
  • Hotel name and reservation details
  • Total price and booking identifier
  • A concise user-facing trip summary

Failure Handling Ideas​

The current checked-in demo focuses on the happy path. Protolink's architecture makes it straightforward to extend the workflow with explicit failure paths:

  • If weather is poor, call the Holiday Advisor for alternate dates or destinations.
  • If hotel booking fails, retry with lower constraints or ask for alternatives.
  • If a tool returns an error part, let the coordinator mark the task as failed or request more input.
  • If the LLM repeats the same action, the inference-loop deduplication guardrail injects corrective feedback.

Extending the Example​

  • Add a flights or ferries agent with a book_transport tool.
  • Add a calendar agent that writes confirmed bookings to a calendar API.
  • Add a messaging agent that sends the final summary over WhatsApp, email, or Slack.
  • Switch transports to SSE JSON-RPC for streamed progress updates.
  • Add MCP tools for real hotel, travel, or messaging integrations.

See Also​