Skip to main content

Core Concepts

Understanding Radium's fundamental concepts will help you build powerful, composable AI systems. This guide introduces the key ideas that make Radium unique.

Vision: Composable Intelligence Infrastructure​

Radium is evolving toward a composable intelligence infrastructure where AI systems are built from reusable, validated components that can automatically assemble themselves. This vision, inspired by the OpenKor architecture, enables:

  • Component Reusability: Build once, use everywhere
  • Automatic Assembly: Systems compose themselves from available components
  • Continuous Improvement: Self-healing and self-improving ecosystems
  • Global Collaboration: Shared component marketplace

Learn more in our Roadmap.

Agents​

Agents are specialized AI assistants configured for specific tasks. Each agent has:

  • Identity: Name, description, and role
  • Capabilities: Defined by prompts and tools
  • Configuration: Model selection, parameters, and behavior
  • Context: Access to workspace files, memory, and other agents

Agent Types​

  • Specialist Agents: Focused on specific domains (code review, architecture, testing)
  • General Agents: Broad capabilities for diverse tasks
  • Orchestrator Agents: Coordinate multiple agents for complex workflows

Agent Lifecycle​

  1. Creation: Define agent configuration and prompts
  2. Discovery: Radium automatically finds agents in workspace
  3. Execution: Agents run tasks using their configured models
  4. Memory: Results stored for future reference
  5. Learning: Agents improve from feedback and usage patterns

Learn more: Agent Configuration

Orchestration​

Orchestration is Radium's intelligent task routing system that automatically:

  • Analyzes your natural language requests
  • Selects the best agent(s) for the task
  • Coordinates multi-agent workflows
  • Synthesizes results from multiple agents

Instead of manually choosing agents, you describe what you need, and the orchestrator handles the rest.

Orchestration Benefits​

  • Natural Interaction: Type requests naturally without command syntax
  • Intelligent Routing: Automatically finds the right specialist
  • Multi-Agent Coordination: Handles complex workflows automatically
  • Model Agnostic: Works with any AI provider

Learn more: Orchestration Guide

Components & Extensions​

Components are reusable building blocks that can be shared and composed:

  • Prompts: Agent prompt templates
  • MCP Servers: Model Context Protocol integrations
  • Commands: Custom CLI commands
  • Hooks: Native or WASM modules for custom behavior

Extensions package components for distribution and sharing.

Component Foundry Pattern​

The Component Foundry Pattern (from OpenKor) provides:

  • Standardized Interfaces: Consistent component patterns
  • Validation Framework: Automated quality checks
  • Composition Rules: Clear guidelines for combining components
  • Version Management: Semantic versioning and compatibility

Learn more: Extension System | Roadmap: Component Foundry

Policies & Security​

Policies provide fine-grained control over agent behavior:

  • Tool Execution Control: Allow, deny, or require approval for specific tools
  • Context-Aware Rules: Different policies for different contexts
  • Approval Modes: Yolo, AutoEdit, or Ask modes
  • Session Constitutions: Temporary rules for specific sessions

Policy Engine​

The policy engine ensures:

  • Safety: Prevent unwanted operations
  • Compliance: Enforce organizational rules
  • Flexibility: Different policies for different scenarios
  • Transparency: Clear policy application and logging

Learn more: Policy Engine

Memory & Context​

Memory enables agents to maintain continuity across sessions:

  • Plan-Scoped Memory: Storage per requirement/plan
  • Agent Output Storage: Automatic persistence of agent results
  • Context Retrieval: Access previous outputs for context

Context Sources provide information to agents:

  • File Sources: Project files and documentation
  • HTTP Sources: External APIs and documentation
  • Jira Integration: Issue tracking integration
  • BrainGrid Integration: Requirement management

Learn more: Memory & Context

Planning & Execution​

Planning converts high-level goals into structured, executable workflows:

  • Goal Decomposition: Break down complex goals into tasks
  • Dependency Analysis: Build dependency graphs (DAGs)
  • Validation: Multi-stage validation with retry logic
  • Workflow Generation: Create executable workflow templates

Execution runs plans with:

  • Automatic Agent Selection: Choose agents for each task
  • Dependency Resolution: Execute tasks in correct order
  • Error Handling: Graceful failure and recovery
  • Progress Tracking: Monitor execution status

Learn more: Autonomous Planning

Persona System​

The Persona System provides intelligent model selection:

  • Cost Optimization: Automatically choose cost-effective models
  • Performance Profiles: Balance speed, cost, and quality
  • Fallback Chains: Automatic fallback to alternative models
  • Model Selection: Primary, fallback, and premium model tiers

Learn more: Persona System

Learning System​

The Learning System tracks and applies knowledge:

  • Mistake Tracking: Learn from errors
  • Preference Learning: Remember user preferences
  • Success Patterns: Identify what works
  • ACE Skillbook: Reusable strategies from past work

Learn more: Learning System

Vibe Check (Metacognitive Oversight)​

Vibe Check provides Chain-Pattern Interrupt (CPI) functionality:

  • Reasoning Lock-In Prevention: Detect when agents get stuck
  • Risk Assessment: Identify potential issues early
  • Pattern Detection: Recognize problematic patterns
  • Phase-Aware Feedback: Adapt to planning, implementation, or review phases

Research shows CPI systems improve success rates by +27% and reduce harmful actions by -41%.

Learn more: Vibe Check

Future: Component Ecosystem​

Radium is evolving toward a global component ecosystem:

Component Foundry​

  • Systematic component creation and validation
  • Quality assurance frameworks
  • Reusable component patterns

Global Component Graph​

  • Discover components across the ecosystem
  • Automatic composition from available components
  • Component relationship tracking

Autonomous Assembly​

  • Systems that compose themselves
  • Goal-driven component selection
  • Dynamic reconfiguration

Learn more: Roadmap: Vision & Innovation

Key Architectural Patterns​

Durable Autonomous Continuous Remediation (DACR)​

Self-healing systems that maintain component quality over time without manual intervention.

Durable Recursive Component Generation (DRCG)​

Components that generate other components recursively, creating self-extending systems.

Autonomous Component-Centric Assembly (ACCA)​

Systems that automatically assemble themselves from available components based on goals and constraints.

Learn more: Roadmap: Vision & Innovation

Next Steps​

Understanding these concepts will help you build more powerful and composable AI systems with Radium.