Agent Jido: the Weather Agent example

Jido is an agent framework built in Elixir, and this post walks through its first worked example: a weather agent that takes a natural language question, decides whether to call a weather tool, and answers in character as an “enthusiastic weather reporter.”

Actions are typed, schema-validated units of functionality, such as Jido.Tools.Weather, which wraps the OpenWeatherMap API; an AI Skill routes incoming messages by signal type and renders them into a prompt template before an Action runs; the Agent process holds state and exposes the public API that ties both together.

Fichamento

#

• Jido.Action validates parameters against a declared schema automatically, and its to_tool/0 function converts the Action’s definition into a JSON schema an LLM framework like Langchain can call as a tool.
• Every user message becomes a Jido.Signal, Jido’s standard messaging format, wrapping the message with a type like "jido.ai.tool.response" before it reaches the agent process.
• The Jido.AI.Skill’s router maps signal types to specific Actions, and its handle_signal/2 callback renders the user’s message into an EEx prompt template before the LLM ever sees it.
• The actual LLM call happens through Jido.AI.Actions.Langchain, which converts Jido Actions into Langchain’s tool-calling format and passes the model, rendered prompt, and available tools together.
• A full request traces through eleven steps: the agent module delegates to a signal, the skill’s router picks the right Action, Langchain calls the LLM, the LLM decides to invoke the weather tool with parsed parameters (e.g. location: "Tokyo"), the tool runs and returns data, and the LLM formats a final response that bubbles back to the caller.
• The example ships with test data mode (test: true) built into the Weather Action’s schema, so it runs end to end in a Livebook without an OpenWeatherMap API key.