Production-grade AI Agent runtime on Elixir/OTP with 8 explicit reasoning strategies and a unified tool contract for building observable, recoverable LLM orchestration systems.
Jido.AI is the AI runtime layer of the Jido ecosystem, built on Elixir/OTP, focused on production-grade LLM Agent orchestration and reasoning control.
Reasoning Strategies#
Defaults to the ReAct (Reasoning + Acting) strategy with 8 switchable explicit reasoning strategies: CoD (Concise Description), CoT (Chain-of-Thought), AoT (Algorithm of Thought), ToT (Tree of Thought), GoT (Graph of Thought), TRM (Thoughtful Recursive Refinement), and Adaptive. Each strategy is implemented as an independent module (e.g., CoTAgent, ToTAgent) sharing a unified Agent interface, selectable by task complexity and latency requirements.
Tool System#
Tools are defined via the Jido.Action macro with compile-time JSON Schema validation using Zoi. Tools can be mounted on Agents or run independently via Jido.Exec.run/3. Request-level control includes allowed_tools for name-based filtering, tool_context for context injection, and request_transformer for dynamic request modification.
Invocation Patterns#
Covers the full spectrum from lightweight to heavyweight: Agent-free shortcuts (generate_text/2, ask/2); synchronous ask_sync/3; asynchronous explicit handles via ask/3 + await/2 preventing concurrent overwrites; and long-lived Agents via AgentServer.
Runtime Guarantees#
Each Agent runs in an isolated BEAM process supervised by OTP Supervisor for fault recovery. Jido.AI.Observe exposes standardized Telemetry events for monitoring integration. The standalone ReAct runtime supports streaming output and checkpoint recovery. Policy and Quota plugins can deterministically intercept unsafe or over-budget requests in the pipeline.
Multi-Provider Support#
Abstracts LLM providers (Anthropic, OpenAI, Google) via req_llm, using semantic aliases (:fast, :capable) for model routing, decoupling business code from specific models.
Jido Ecosystem Layers#
jido_ai(this project): AI runtime layer — reasoning strategies, Agent DSL, tool orchestrationjido_action: Typed tool/Action contract (schema + run callback)jido: Agent runtime, process model, Signal lifecyclereq_llm: LLM provider abstraction layer (Anthropic/OpenAI/Google)
Installation#
Recommended via Igniter:
mix igniter.install jido_ai
Manual installation requires adding {:jido, "~> 2.0"} and {:jido_ai, "~> 2.1.0"} to mix.exs dependencies, then configuring model_aliases and provider API keys.