It shows a small collaborative agent team where a root chat-mode coordinator delegates work to two specialist sub-agents — one in single_turn mode and one in task mode — and they each return control automatically when done.

The three modes work together:

For mode semantics see Build collaborative agent teams.

• What's the weather like at home today?
• Is it raining in Tokyo?
• Please book me a flight from Zurich to London on 2026-07-12. I'm Alex Müller, window seat.
• Book a flight Zurich → Barcelona on 2026-08-01. (the task agent will ask for the missing passenger name)

graph TD
    travel_planner -->|delegates| weather_checker
    travel_planner -->|delegates| flight_booker
    weather_checker -.->|uses| user_info[user_info tool]
    weather_checker -.->|uses| geocode_address[geocode_address tool]
    weather_checker -.->|uses| get_weather[get_weather tool]
    flight_booker -.->|uses| search_flights[search_flights tool]
    flight_booker -.->|uses| book_flight[book_flight tool]

1. Build a single_turn sub-agent for autonomous lookups:

   weatherChecker, err := llmagent.New(llmagent.Config{
       Name:  "weather_checker",
       Model: model,
       Mode:  llmagent.ModeSingleTurn,
       Tools: []tool.Tool{getWeatherTool, userInfoTool, geocodeAddressTool},
       // ...
   })

2. Build a task sub-agent for structured, possibly multi-turn data work:

   flightBooker, err := llmagent.New(llmagent.Config{
       Name:         "flight_booker",
       Model:        model,
       Mode:         llmagent.ModeTask,
       InputSchema:  flightInputSchema,
       OutputSchema: flightResultSchema,
       Tools:        []tool.Tool{searchFlightsTool, bookFlightTool},
       // ...
   })

3. Assign both to a coordinator (chat is the default mode for a sub-agent):

   travelPlanner, err := llmagent.New(llmagent.Config{
       Name:      "travel_planner",
       Model:     model,
       SubAgents: []agent.Agent{weatherChecker, flightBooker},
       // ...
   })

The coordinator delegates by mentioning the sub-agent's name in its instruction. ADK auto-injects the matching delegation tool for each sub-agent. A single_turn or task sub-agent returns control to the coordinator automatically once it produces its result; no explicit transfer_to_agent call is required (this contrasts with default chat mode, which requires a manual handoff).

Set your API key and run the sample with the console launcher:

export GOOGLE_API_KEY=...
go run ./examples/multiagent/collaboration console

You can also serve it over REST or the web UI; run with help for all launcher subcommands:

go run ./examples/multiagent/collaboration help

• examples/multiagent/single_turn — a single_turn sub-agent in isolation.
• examples/multiagent/task_sub_agent — two task sub-agents in isolation.