Build agents with Elixir

ExAgent · Elixir / OTP

Start with a model and a tool.
Keep control as your application grows.

One agent loop, with optional layers for conversations, persistence, approvals and coordination. Stock ReqLLM connects your models; your application owns its data and effects.

Run your first agent → Integrate with a coding agent

Using ExAgent 2.1

These pages describe ExAgent 2.1.0; published 1.x packages have older contracts. Check supported features and limits, use the dependency in Getting started, or read Migration when upgrading an existing application.

A small, working beginning

Run this inside an application that depends on ExAgent2.0. No credentials or network are needed:

agent = ExAgent.new(model: "test", instructions: "Be concise.")
{:ok, result} = ExAgent.run(agent, "Hello!")
result.output
# => "a test response"

The test model follows the same loop as a provider. Next, add a tool, validate a typed result, or replace the model with a qualified ReqLLM instance.

Choose the layer your application needs

Run: A request, tools and a validated result. Start here for scripts, jobs and extraction.
Server: A supervised conversation with history, a bounded queue and events.
Store: Confirmed snapshots and explicit continuation boundaries. Add durability when you need it.
Coordination: Session turns, scoped delegation, durable sequences and bounded router or parallel flows.

Each layer has its own ownership contract. A checkpoint confirms saved data; an application still reconciles an external effect whose outcome is uncertain.

Build something useful

Your next task	Guide
Call application code and return validated data	Tools and structured output
Select a provider and bound work	Models, budgets and limits
Add conversation history, streaming and UI events	Runtime and events
Persist a conversation or wait for human approval	Durability and approvals
Coordinate several agents	Coordination
Test without calling a model provider	Testing
Export traces to Langfuse or Opik	Observability
Connect an external tool server	MCP

For an unexpected result, start with Troubleshooting.

Read the contract behind the example

ExAgent.run/3 is the main entry point. ExAgent.Tool describes a tool; ExAgent.Server owns a conversation. Browse the Modules tab or use / to search functions and pages. Guides link to the relevant API as you go.

For coding agents, Integration notes provide an API map, result shapes, responsibilities and a reading route. ExDoc also generates llms.txt and Markdown versions of these pages from the same source. Use the View llms.txt action at the bottom of the generated page.

Supported features and limits explains deployment boundaries. The documentation map connects tutorials, integration recipes and the API reference.

Next Page → Getting started