Core Concepts

This page explains the fundamental concepts that power FAOS. Understanding these concepts will help you get the most out of the platform.

Agents

Agents are autonomous AI entities that perform work on your behalf. Unlike simple chatbots, FAOS agents can:

• Execute multi-step workflows
• Make decisions within defined guardrails
• Learn from interactions
• Coordinate with other agents

Agent Anatomy

┌─────────────────────────────────────┐
│             FAOS Agent              │
├─────────────────────────────────────┤
│  🧠 LLM Core (Claude/GPT/etc.)      │
│  📚 Industry Ontology               │
│  🔧 Tool Access (APIs, Integrations)│
│  🛡️ Guardrails & Policies          │
│  💾 Memory (Context & History)      │
└─────────────────────────────────────┘

Ontologies

Ontologies are structured knowledge frameworks that give agents industry-specific context. They include:

• Terminology: Domain-specific vocabulary
• Entities: Key objects (customers, orders, claims, etc.)
• Workflows: Common business processes
• Regulations: Compliance requirements
• Prompt Templates: Pre-built patterns for common tasks

Example: Banking Ontology

entities:
  - Account (Checking, Savings, Investment)
  - Transaction (Deposit, Withdrawal, Transfer)
  - Customer (Individual, Business)

workflows:
  - Account Opening
  - Fraud Detection
  - Loan Processing

regulations:
  - KYC (Know Your Customer)
  - AML (Anti-Money Laundering)
  - PCI-DSS (Payment Card Security)

Graduated Autonomy

FAOS implements a Graduated Autonomy model that lets you control how independently agents operate.

Level	Name	Description
L1	Assisted	AI suggests actions, human executes
L2	Supervised	AI executes, human approves first
L3	Autonomous	AI executes within guardrails
L4	Proactive	AI anticipates needs and acts

Autonomy Progression

Agents can earn higher autonomy levels based on:

• Accuracy of past decisions
• User approval rates
• Compliance with guardrails
• Time since last escalation

Guardrails

Guardrails are safety constraints that define what agents can and cannot do.

Types of Guardrails

1. Action Guardrails: What actions require approval
2. Data Guardrails: What data can be accessed
3. Rate Guardrails: How many actions per time period
4. Escalation Guardrails: When to involve humans

const guardrails = {
  actions: {
    requireApproval: ["delete", "refund", "modify-account"],
    blocked: ["share-pii-externally"]
  },
  data: {
    canAccess: ["orders", "products", "public-profiles"],
    cannotAccess: ["payment-details", "ssn"]
  },
  rates: {
    maxActionsPerMinute: 10,
    maxActionsPerHour: 100
  }
};

Workflows

Workflows are sequences of actions that agents execute. They can be:

• Pre-built: Industry templates from ontologies
• Custom: Defined by your organization
• Learned: Discovered from user behavior

Workflow Example

graph LR
    A[Customer Request] --> B{Classify Intent}
    B -->|Refund| C[Check Policy]
    B -->|Question| D[Search KB]
    C --> E{Amount > $100?}
    E -->|Yes| F[Request Approval]
    E -->|No| G[Process Refund]
    D --> H[Generate Response]

Memory

Agents maintain memory across interactions:

• Short-term: Current conversation context
• Long-term: Historical interactions with entities
• Shared: Knowledge shared across agents

Multi-Agent Coordination

Multiple agents can work together on complex tasks:

• Handoffs: Transfer conversations between specialized agents
• Parallel Processing: Divide tasks among agents
• Consensus: Require agreement before critical actions

---

Next Steps

• Architecture Overview - Technical deep-dive
• User Guides - Practical walkthroughs
• API Reference - Developer documentation