← back to ecosystem
ai stable version 2.0.0 hex 2.0.0 package jido_ai elixir ~> 1.17

Jido AI

Jido AI turns raw LLM calls into structured agent intelligence with strategy-driven reasoning, tool use, and accuracy controls.

Stable support. Ongoing maintenance, compatibility work, and careful API evolution.

Docs Hex GitHub ★179 Issues
View package metadata source →

HOW MODULES FIT TOGETHER

Interface
Reasoning Core
Model + Tools
ReqLLM

Provider adapters and model invocation

Flow
Ask/Await API -> Strategies (select strategy) Strategies -> Accuracy Pipeline (produce and score candidates) Strategies -> ReqLLM (call models) Strategies -> Tool Execution (invoke tools) Tool Execution -> Action Contract (typed actions) Ask/Await API -> Agent Runtime (run inside agents)

RELATED PACKAGES

AT A GLANCE

Start with a single ask/await workflow, then scale to orchestrated multi-agent reasoning.
Reasoning strategies (ReAct, CoT, ToT, GoT, TRM, Adaptive) let you tune cost vs quality.
Accuracy pipeline layers verification, reflection, and self-consistency on top of model outputs.
Integrates directly with jido runtime, req_llm providers, and jido_browser tool execution.
Designed for production reliability with async tracking, model aliases, and signal-driven orchestration.

DEEP DIVE

Overview

Jido AI is the AI integration layer for the Jido ecosystem, providing LLM orchestration, reasoning strategies, and research-backed accuracy improvement techniques for building intelligent agents in Elixir. Built on top of ReqLLM for multi-provider LLM access, it turns raw LLM calls into structured, reliable agent behaviors through composable strategies, accuracy pipelines, and a robust tool execution framework.

Purpose

Jido AI is the general AI strategy/action ecosystem adjacent to the CLI-agent runtime stack.

Boundary Lines

  • Owns LLM reasoning, strategy selection, and accuracy improvement capabilities.
  • Provides reusable intelligence primitives for higher-level orchestration packages.
  • Does not own CLI provider adapter parity, runtime preflight policy, or shell/filesystem substrate concerns.

Major Components

Reasoning Strategies

Six pluggable strategies: ReAct (tool-using agents), Chain-of-Thought, Tree-of-Thoughts, Graph-of-Thoughts, TRM (Tiny-Recursive-Model with adaptive computation), and Adaptive (automatic strategy selection).

Accuracy Improvement Pipeline

Multi-stage pipeline with self-consistency voting, search algorithms (Beam Search, MCTS), verifiers (LLM, code execution, deterministic, static analysis, unit test), difficulty estimation, reflection/revision loops, and confidence calibration.

Pre-Built Actions

Composable Jido Actions for LLM operations (Chat, Complete, Embed, GenerateObject), Reasoning (Analyze, Explain, Infer), Planning (Plan, Decompose, Prioritize), Orchestration (DelegateTask, SpawnChildAgent), Streaming, and Tool Calling.

Agent Macros

Ready-to-use agent bases: ReActAgent, CoTAgent, ToTAgent, GoTAgent, TRMAgent, AdaptiveAgent, and OrchestratorAgent.

Plugins

Composable skill bundles for LLM, Reasoning, Planning, Orchestration, Streaming, Tool Calling, and TaskSupervisor capabilities.