Package Overview (Production Map)
View SourceYou want a clear map of what jido_ai provides before production rollout.
After this guide, you should be able to explain the package in a prioritized way and choose the right integration surface for each workload.
Core Product Surface
jido_ai is an AI runtime layer for Jido agents with:
- Agent macros with built-in reasoning strategies
- Tool-calling orchestration over
Jido.Actionmodules - Strategy-agnostic skills loading and registry
- Plugin-based capability mixins for reusable runtime features
- Action APIs for direct AI workflows outside agent macros
- First-class observability via signals, directives, and telemetry
Priority 1: Jido.AI.Agent (Default ReAct Agent)
Jido.AI.Agent is the anchor feature.
- It is the generic AI agent macro.
- Under the hood it uses
Jido.AI.Reasoning.ReAct.Strategy. - It is built for tool use through a ReAct loop (reason, call tool, continue).
- It supports request handles and async orchestration (
ask/await/ask_sync). - ReAct requests can narrow tools per run with
allowed_toolsor fully override them withtools. - It uses standardized lifecycle and runtime contracts (
ai.request.*,ai.llm.*,ai.tool.*).
If your production workload needs reliable tool-calling agents, this is the default entry point.
Priority 2: Multi-Strategy Agent Macros
jido_ai ships specialized agent macros for different reasoning patterns:
Jido.AI.CoDAgent->Jido.AI.Reasoning.ChainOfDraft.StrategyJido.AI.CoTAgent->Jido.AI.Reasoning.ChainOfThought.StrategyJido.AI.AoTAgent->Jido.AI.Reasoning.AlgorithmOfThoughts.StrategyJido.AI.ToTAgent->Jido.AI.Reasoning.TreeOfThoughts.StrategyJido.AI.GoTAgent->Jido.AI.Reasoning.GraphOfThoughts.StrategyJido.AI.TRMAgent->Jido.AI.Reasoning.TRM.StrategyJido.AI.AdaptiveAgent->Jido.AI.Reasoning.Adaptive.Strategy
Supported strategy family:
- ReAct
- Chain-of-Draft
- Chain-of-Thought
- Algorithm-of-Thoughts
- Tree-of-Thoughts
- Graph-of-Thoughts
- TRM
- Adaptive
Use these when reasoning policy is part of the product requirement, not just a model prompt detail.
Tree-of-Thoughts Runtime Contract (Production)
Jido.AI.ToTAgent now returns a structured result contract (hard switch), not a plain string:
best: best-ranked candidatecandidates: top-K ranked leavestermination: reason/status/depth/node-count/durationtree: traversal and search-shape metadatausage: accumulated token usagediagnostics: parser mode/retries, convergence, tool-round diagnostics
ToT flexibility controls exposed at the agent macro level:
top_k,min_depth,max_nodes,max_duration_ms,beam_widthearly_success_threshold,convergence_window,min_score_improvementmax_parse_retriestools,tool_context,request_transformer,tool_timeout_ms,tool_max_retries,tool_retry_backoff_ms,max_tool_round_trips
Priority 3: Skills System (SKILL.md / skills.io-aligned workflow)
Skills are reusable instruction/capability units loaded at runtime:
Jido.AI.Skill.Loaderparses skill files.Jido.AI.Skill.Registrystores and resolves active skills.- Skills are strategy-agnostic and can be injected into agent behavior/context.
- Skills are useful for domain behavior reuse without duplicating agent code.
This is the main packaging mechanism for reusable AI behavior in your system.
Priority 4: Plugins As Capability Mixins
Plugins should represent product capabilities, not low-level runtime plumbing.
Recommended plugin set (target production surface):
Jido.AI.Plugins.Chat(anchor capability)- Unified conversational interface with built-in tool calling.
- Replaces split end-user mental model of separate
LLM+ToolCallingplugins. - Supports simple chat usage and tool-augmented chat under one contract.
Jido.AI.Plugins.Planning- Structured planning, decomposition, and prioritization flows.
- Strategy invocation plugins (explicit reasoning as capabilities):
Jido.AI.Plugins.Reasoning.ChainOfDraftJido.AI.Plugins.Reasoning.ChainOfThoughtJido.AI.Plugins.Reasoning.AlgorithmOfThoughtsJido.AI.Plugins.Reasoning.TreeOfThoughtsJido.AI.Plugins.Reasoning.GraphOfThoughtsJido.AI.Plugins.Reasoning.TRMJido.AI.Plugins.Reasoning.Adaptive- These expose reasoning strategies as callable plugin capabilities, independent of which agent macro strategy is primary.
- Cross-cutting policy plugins (production hardening):
- model routing/fallback
- guardrails and safety policy
- memory/retrieval enrichment
- cost/quota/rate limiting
Non-goal for public plugin surface:
TaskSupervisorshould not be treated as a primary user-facing capability plugin.- Async execution infrastructure should be handled by Jido runtime/Jido.Exec primitives and internal runtime wiring.
Where plugins fit:
- Strategies decide control flow and reasoning mechanics.
- Plugins package reusable capabilities and policy.
- Actions remain the executable units behind both plugins and direct workflows.
Signal namespace contract for this plugin surface:
chat.messagereasoning.cod.runreasoning.cot.runreasoning.aot.runreasoning.tot.runreasoning.got.runreasoning.trm.runreasoning.adaptive.run
Priority 5: Actions As Independent Integration Surface
Standalone actions are the strategy-independent integration path for adding AI behavior directly to Jido apps via Jido.Exec.
Finalized standalone action set (recommended):
- Core generation primitives
Jido.AI.Actions.LLM.Chat(single-turn conversational generation)Jido.AI.Actions.LLM.GenerateObject(schema-constrained structured output)Jido.AI.Actions.LLM.Embed(embedding generation for retrieval/search)
- Tool orchestration primitives
Jido.AI.Actions.ToolCalling.CallWithTools(LLM + tool schema + optional auto-execution loop)Jido.AI.Actions.ToolCalling.ExecuteTool(direct tool execution by name)Jido.AI.Actions.ToolCalling.ListTools(tool discovery and schema inspection)
- Planning domain templates
- Reasoning domain templates (optional, useful outside full strategy orchestration)
- Dedicated strategy orchestration
Jido.AI.Actions.Reasoning.RunStrategy(isolated strategy execution for:cod | :cot | :aot | :tot | :got | :trm | :adaptive)
- Compatibility convenience
Jido.AI.Actions.LLM.Complete(simple completion path; overlaps withChatand can remain as convenience)
Not part of standalone action surface:
- Strategy-internal command actions (
*_start,*_llm_result,*_llm_partial, worker lifecycle events). - These are orchestration internals for ReAct/CoD/CoT/AoT/ToT/GoT/TRM/Adaptive strategies, not reusable app-level primitives.
Pragmatically:
- Agent macros are the primary production surface for long-lived agent orchestration.
- Direct actions are the flexible lower-level surface for pipelines, jobs, and custom runtime composition.
Runtime Map (End-To-End)
User/App Query
-> Agent Macro (Jido.AI.Agent or strategy-specific agent)
-> Strategy (ReAct/CoD/CoT/AoT/ToT/GoT/TRM/Adaptive)
-> Directives (LLM, tool, control intents)
-> Runtime Execution (ReqLLM + tool execution)
-> Signals (ai.request.*, ai.llm.*, ai.tool.*, ai.usage)
-> Strategy state updates
-> Request completion/await resultObservability Guarantees
Observability is a core part of the package, not an add-on:
- Typed signal contracts for lifecycle, LLM, tool, and usage events
- Directive-level execution boundaries
- Telemetry events for request, LLM, and tool phases
- Request IDs and run IDs for correlation across async boundaries
Production Positioning Summary
When describing jido_ai for production, the concise version is:
- A ReAct-first AI agent framework (
Jido.AI.Agent) with built-in tool calling. - A multi-strategy reasoning platform (CoD, CoT, AoT, ToT, GoT, TRM, Adaptive, ReAct).
- A reusable skills layer for domain behavior packaging.
- A plugin layer for mountable capability mixins and policy controls.
- A lower-level actions API for strategy-independent AI workflows.
- Full runtime observability through directives, signals, and telemetry.