AI Agents Overview

Understanding the different types of AI agents in Fabric and when to use each one.

Fabric AI includes multiple specialized agents, each designed for specific tasks. This guide helps you understand which agent to use and how they work together.

What is an Agent?

An AI agent is an autonomous system that can:

Understand — Interpret your requests and context
Plan — Break down complex tasks into steps
Execute — Perform actions like generating documents or calling APIs
Learn — Improve from feedback and past executions

Unlike simple chatbots, agents can take actions, use tools, and complete multi-step workflows.

Agent Types

Fabric provides several types of agents:

┌─────────────────────────────────────────────────────────────────┐
│                        Fabric AI Agents                          │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  ┌──────────────────────────────────────────────────────────┐   │
│  │                    FABRIC ORCHESTRATOR                    │   │
│  │         Routes tasks to the right specialized agent       │   │
│  └──────────────────────────────────────────────────────────┘   │
│                              │                                   │
│          ┌───────────────────┼───────────────────┐              │
│          │                   │                   │              │
│          ▼                   ▼                   ▼              │
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐      │
│  │   Document   │    │     CUGA     │    │    Custom    │      │
│  │  Generator   │    │ (General AI) │    │    Agents    │      │
│  └──────────────┘    └──────────────┘    └──────────────┘      │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Fabric Orchestrator

The intelligent coordinator that routes your requests to the right agent or tool.

Best for:

Complex multi-step tasks
Tasks requiring multiple tools
When you're not sure which agent to use
Workflow automation

Capabilities:

Semantic routing to find the best executor
Task decomposition and planning
MCP tool execution
Agent delegation via A2A protocol
Trust-based approval workflows
Memory and learning from past executions

Learn more about the Orchestrator →

Document Generator

A specialized agent for creating structured documents.

Best for:

PRDs and product documentation
Technical specifications
Architecture documents
User stories and acceptance criteria
API documentation

Capabilities:

Multiple document templates
RAG context integration
Real-time streaming with diff highlighting
Confirm/reject workflow for changes
Export to PDF, DOCX, Markdown

Learn more about Document Generation →

CUGA (Configurable Generalist Agent)

A powerful generalist agent for complex automation.

Best for:

Browser automation and web scraping
Complex API orchestration
Multi-step workflows
Tasks requiring code execution
Autonomous task completion

Capabilities:

Full Playwright browser automation
Sandboxed code execution
Screenshot analysis
Human-in-the-loop approvals
Memory system for learning

Learn more about CUGA →

Custom Agents

Create your own agents for specific workflows.

Best for:

Organization-specific processes
Unique document types
Custom integrations
Specialized domains

Capabilities:

Custom prompts and instructions
Configurable MCP tools
Custom approval workflows
Template-based creation

Learn how to create custom agents →

Choosing the Right Agent

Task	Recommended Agent	Why
Generate a PRD	Document Generator	Optimized for structured documents
Create 10 Jira tickets from PRD	Orchestrator	Multi-step with MCP tools
Scrape competitor websites	CUGA	Browser automation needed
Generate weekly report	Custom Agent	Organization-specific format
"Just help me with X"	Orchestrator	Routes automatically

How Agents Work

1. Request Understanding

When you send a message:

User: "Create a PRD for user authentication with OAuth support"
         │
         ▼
┌─────────────────────────────────────────────────────────┐
│                  Request Analysis                        │
│  • Task type: Document generation                        │
│  • Document type: PRD                                    │
│  • Key topics: Authentication, OAuth                     │
│  • Complexity: Medium                                    │
└─────────────────────────────────────────────────────────┘

2. Context Retrieval

The agent gathers relevant context:

┌─────────────────────────────────────────────────────────┐
│                  Context Sources                         │
│  • Project workspace documents                           │
│  • Previous project documents                            │
│  • Organization knowledge base                           │
│  • Past similar requests                                 │
└─────────────────────────────────────────────────────────┘

3. Planning

For complex tasks, the agent creates a plan:

┌─────────────────────────────────────────────────────────┐
│                    Execution Plan                        │
│  Step 1: Retrieve auth-related documents                 │
│  Step 2: Generate PRD outline                            │
│  Step 3: Expand each section                             │
│  Step 4: Add acceptance criteria                         │
│  Step 5: Review and format                               │
└─────────────────────────────────────────────────────────┘

4. Execution

The agent executes each step:

AI generation — Call LLM to generate content
Tool calls — Execute MCP tools if needed
Validation — Check output quality
Streaming — Send updates to user in real-time

5. Learning

After completion:

Record what worked and what didn't
Update memory for future requests
Improve routing accuracy

Agent Communication

A2A Protocol

Agents communicate using the Agent-to-Agent (A2A) protocol:

Orchestrator                    Document Generator
     │                                  │
     │─── Discover capabilities ───────>│
     │<── Agent manifest ───────────────│
     │                                  │
     │─── Task: Generate PRD ──────────>│
     │<── SSE: Progress updates ────────│
     │<── SSE: Content chunks ──────────│
     │<── Result: Complete document ────│
     │                                  │

AG-UI Protocol

Agents communicate with the UI using the AG-UI protocol:

State deltas — Real-time state updates
Tool calls — Visibility into agent actions
Streaming — Token-by-token generation
Approvals — Human-in-the-loop requests

Multi-Agent Workflows

The Orchestrator can coordinate multiple agents:

Request: "Create a PRD, then generate Jira stories"
                    │
                    ▼
        ┌───────────────────┐
        │   Orchestrator    │
        │   (Coordinator)   │
        └───────────────────┘
                    │
        ┌───────────┴───────────┐
        │                       │
        ▼                       ▼
┌──────────────┐       ┌──────────────┐
│  Document    │       │   Jira MCP   │
│  Generator   │──────>│     Tools    │
└──────────────┘       └──────────────┘
   (Step 1)               (Step 2)
                             │
                             ▼
                    Stories created in Jira