Building a Weather Agent

Jump To Section

Setup Configure credentials Define the weather Actions Inspect the generated tool definitions Add a custom tool action Define the weather agent Start the runtime and agent Ask for a forecast Ask a follow-up with ask/3 and await/2 How tool-calling works

Setup

This notebook is self-contained. It defines the weather Actions locally, configures one provider key, starts the default runtime, and runs a tool-calling weather agent end to end. The weather data itself comes from free public APIs, so you do not need a separate weather API key.

Mix.install([
  {:jido, "~> 2.1"},
  {:jido_ai, "~> 2.0"},
  {:req_llm, "~> 1.7"},
  {:req, "~> 0.5"}
])

Logger.configure(level: :warning)

# Livebook imports can execute generated docs as doctests.
# Disable compiler docs until the current Jido Hex release drops the invalid signal_types/0 example.
Code.put_compiler_option(:docs, false)

Configure credentials

In Livebook, store OPENAI_API_KEY as a secret. Livebook exposes it as LB_OPENAI_API_KEY, so the cell below checks both names.

openai_key = System.get_env("LB_OPENAI_API_KEY") || System.get_env("OPENAI_API_KEY")

configured? =
  if is_binary(openai_key) do
    ReqLLM.put_key(:openai_api_key, openai_key)
    true
  else
    IO.puts("Set OPENAI_API_KEY or LB_OPENAI_API_KEY before running the request cells.")
    false
  end

Define the weather Actions

In Jido, tools are Actions. This notebook defines its weather Actions locally so the example stays runnable in Livebook and in a fresh Mix project.

defmodule MyApp.WeatherHTTP do
  @headers [
    {"user-agent", "jido.run weather guide (https://jido.run)"},
    {"accept", "application/geo+json, application/json"}
  ]

  def get_json(url, opts \\ []) do
    with {:ok, response} <-
           Req.get(
             url: url,
             params: Keyword.get(opts, :params, []),
             headers: @headers
           ) do
      {:ok, response.body}
    end
  end
end

defmodule MyApp.WeatherGeocode do
  use Jido.Action,
    name: "weather_geocode",
    description: "Convert a city or location string into latitude and longitude coordinates.",
    schema: [
      location: [type: :string, required: true, doc: "City, state, or other human-readable location"]
    ]

  @impl true
  def run(%{location: location}, _context) do
    with {:ok, results} <-
           MyApp.WeatherHTTP.get_json(
             "https://nominatim.openstreetmap.org/search",
             params: [q: location, format: "jsonv2", limit: 1]
           ),
         [match | _] <- results,
         {latitude, _} <- Float.parse(match["lat"]),
         {longitude, _} <- Float.parse(match["lon"]) do
      {:ok,
       %{
         location: location,
         display_name: match["display_name"],
         latitude: latitude,
         longitude: longitude,
         coordinates: "#{latitude},#{longitude}"
       }}
    else
      [] -> {:error, :location_not_found}
      :error -> {:error, :invalid_coordinates}
      {:error, reason} -> {:error, reason}
    end
  end
end

defmodule MyApp.WeatherLocationToGrid do
  use Jido.Action,
    name: "weather_location_to_grid",
    description: "Resolve a latitude/longitude pair into the National Weather Service forecast URLs.",
    schema: [
      coordinates: [
        type: :string,
        required: true,
        doc: "Latitude and longitude as \"lat,lng\""
      ]
    ]

  @impl true
  def run(%{coordinates: coordinates}, _context) do
    with [latitude, longitude] <- String.split(coordinates, ",", parts: 2),
         latitude = String.trim(latitude),
         longitude = String.trim(longitude),
         {:ok, body} <- MyApp.WeatherHTTP.get_json("https://api.weather.gov/points/#{latitude},#{longitude}"),
         properties when is_map(properties) <- body["properties"] do
      {:ok,
       %{
         coordinates: coordinates,
         forecast_url: properties["forecast"],
         hourly_forecast_url: properties["forecastHourly"],
         observation_stations_url: properties["observationStations"],
         office: properties["gridId"],
         grid_x: properties["gridX"],
         grid_y: properties["gridY"]
       }}
    else
      {:error, reason} -> {:error, reason}
      _ -> {:error, :invalid_coordinates}
    end
  end
end

defmodule MyApp.WeatherForecast do
  use Jido.Action,
    name: "weather_forecast",
    description: "Fetch the upcoming weather forecast from a National Weather Service forecast URL.",
    schema: [
      forecast_url: [type: :string, required: true, doc: "Forecast URL returned by weather_location_to_grid"]
    ]

  @impl true
  def run(%{forecast_url: forecast_url}, _context) do
    with {:ok, body} <- MyApp.WeatherHTTP.get_json(forecast_url),
         periods when is_list(periods) <- get_in(body, ["properties", "periods"]) do
      preview =
        periods
        |> Enum.take(4)
        |> Enum.map(fn period ->
          %{
            name: period["name"],
            temperature: period["temperature"],
            temperature_unit: period["temperatureUnit"],
            short_forecast: period["shortForecast"],
            detailed_forecast: period["detailedForecast"]
          }
        end)

      summary =
        Enum.map_join(preview, "\n", fn period ->
          "#{period.name}: #{period.temperature}#{period.temperature_unit}, #{period.short_forecast}"
        end)

      {:ok, %{summary: summary, periods: preview}}
    else
      {:error, reason} -> {:error, reason}
      _ -> {:error, :forecast_unavailable}
    end
  end
end

The weather Actions follow a simple chain:

weather_geocode turns a city into coordinates
weather_location_to_grid turns those coordinates into NWS URLs
weather_forecast fetches the forecast from the returned URL

Inspect the generated tool definitions

ToolAdapter converts Actions into the JSON Schema format that the LLM provider expects for tool calling:

tools = Jido.AI.ToolAdapter.from_actions([
  MyApp.WeatherGeocode,
  MyApp.WeatherLocationToGrid,
  MyApp.WeatherForecast
])

IO.inspect(hd(tools).name)
IO.inspect(hd(tools).parameter_schema)

Add a custom tool action

Any Jido.Action module can be exposed to the LLM as a tool. This example adds a simple temperature converter so the agent can answer unit-conversion questions too.

defmodule MyApp.TemperatureConverter do
  use Jido.Action,
    name: "convert_temperature",
    description: "Convert between Fahrenheit and Celsius",
    schema: [
      value: [type: :float, required: true, doc: "Temperature value"],
      from: [
        type: {:in, [:fahrenheit, :celsius]},
        required: true,
        doc: "Source unit"
      ],
      to: [
        type: {:in, [:fahrenheit, :celsius]},
        required: true,
        doc: "Target unit"
      ]
    ]

  @impl true
  def run(%{value: value, from: :fahrenheit, to: :celsius}, _context) do
    {:ok, %{result: Float.round((value - 32) * 5 / 9, 1), unit: "°C"}}
  end

  def run(%{value: value, from: :celsius, to: :fahrenheit}, _context) do
    {:ok, %{result: Float.round(value * 9 / 5 + 32, 1), unit: "°F"}}
  end

  def run(%{value: value, from: same, to: same}, _context) do
    unit = if same == :celsius, do: "°C", else: "°F"
    {:ok, %{result: value, unit: unit}}
  end
end

Define the weather agent

use Jido.AI.Agent wires in the ReAct reasoning strategy, which handles the tool-calling loop automatically.

defmodule MyApp.WeatherAgent do
  use Jido.AI.Agent,
    name: "weather_agent",
    description: "Weather assistant with tool access",
    tools: [
      MyApp.WeatherGeocode,
      MyApp.WeatherLocationToGrid,
      MyApp.WeatherForecast,
      MyApp.TemperatureConverter
    ],
    model: :fast,
    max_iterations: 6,
    system_prompt: """
    You are a helpful weather assistant.

    For weather questions:
    1. Call weather_geocode with the user's location.
    2. Call weather_location_to_grid with the coordinates from weather_geocode.
    3. Call weather_forecast with the forecast_url from weather_location_to_grid.
    4. Summarize the forecast and give practical advice.

    Use convert_temperature only when the user explicitly asks for unit conversion.
    Keep answers concise and practical.
    """

  @default_timeout 60_000

  def get_forecast(pid, location, opts \\ []) do
    ask_sync(
      pid,
      "What's the forecast in #{location}? Summarize the next few forecast periods.",
      Keyword.put_new(opts, :timeout, @default_timeout)
    )
  end

  def need_jacket?(pid, location, opts \\ []) do
    ask_sync(
      pid,
      "Should I bring a jacket in #{location} today? Use the forecast and give practical advice.",
      Keyword.put_new(opts, :timeout, @default_timeout)
    )
  end
end

The key options control agent behavior:

tools exposes Jido.Action modules to the LLM
model: :fast uses the standard fast-model alias from jido_ai
max_iterations caps the number of ReAct loops before the agent stops
system_prompt teaches the model the required tool order

Start the runtime and agent

{:ok, _} = Jido.start()
runtime = Jido.default_instance()
agent_id = "weather-demo-#{System.unique_integer([:positive])}"

{:ok, pid} = Jido.start_agent(runtime, MyApp.WeatherAgent, id: agent_id)

Ask for a forecast

Show the success path first: ask one forecast question and print the answer.

forecast_answer =
  if configured? do
    {:ok, answer} = MyApp.WeatherAgent.get_forecast(pid, "Chicago, IL")
    answer
  else
    "Skipped request. Set OPENAI_API_KEY or LB_OPENAI_API_KEY to run it."
  end

IO.puts(forecast_answer)

Ask a follow-up with `ask/3` and `await/2`

For non-blocking usage, ask/3 returns a request handle you can await later:

jacket_answer =
  if configured? do
    {:ok, request} =
      MyApp.WeatherAgent.ask(
        pid,
        "Should I bring a jacket in Denver, CO today?"
      )

    {:ok, answer} = MyApp.WeatherAgent.await(request, timeout: 60_000)
    answer
  else
    "Skipped request. Set OPENAI_API_KEY or LB_OPENAI_API_KEY to run it."
  end

IO.puts(jacket_answer)

How tool-calling works

The agent uses a ReAct loop to decide when tools are needed:

Your query is sent to the LLM with JSON Schema definitions of the available tools
The LLM either answers directly or requests a tool call
Jido executes the matching Action’s run/2
The tool result is sent back to the LLM as context
Steps 2 through 4 repeat until the LLM gives a final answer or max_iterations is reached

For a weather question like “What’s the forecast in Chicago?”, the loop usually runs three tool steps: geocode, points lookup, and forecast retrieval.

Error handling

ask_sync/3 returns {:error, reason} for provider timeouts, API failures, and max iteration exhaustion. Pattern match on the result so callers can handle failures explicitly.

query = "What's the forecast in Denver, CO? Do I need a jacket?"

case if(configured?, do: MyApp.WeatherAgent.ask_sync(pid, query, timeout: 60_000), else: {:skip, :no_openai_key}) do
  {:ok, answer} ->
    IO.puts(answer)

  {:error, reason} ->
    IO.puts("Failed: #{inspect(reason)}")

  {:skip, :no_openai_key} ->
    IO.puts("Skipped request. Set OPENAI_API_KEY or LB_OPENAI_API_KEY to run it.")
end

Common failure modes include:

Provider rate limits or network errors from the LLM API
Geocoding returning no results for the requested location
Weather endpoint errors from Nominatim or the NWS API
Hitting max_iterations without reaching a final answer

Next steps

Continue to Build an AI Chat Agent for multi-turn conversation on the same agent process.
Continue to Testing agents and actions to make your custom Actions deterministic under test.
Continue to AI agent with tools when you want the lower-level tool-calling concepts explained step by step.

Edit this page →