Flezi AgentBox

Builder Lab Guide

Complete guide to building AI agents with the three-phase Builder Lab workflow

Builder Lab Guide

The Builder Lab is Flezi AgentBox's agent creation environment. It guides you through a structured three-phase workflow to build, configure, and publish AI agents.

Overview

The Builder Lab uses a phased approach to agent creation:

  1. Discovery Phase — Define your agent's purpose, knowledge base, and capabilities
  2. Logic Phase — Design the processing pipeline with a visual DAG editor
  3. Synthesis Phase — Generate the final blueprint and publish

Each phase builds on the previous one, ensuring your agent has a solid foundation before you define its behavior.

Getting Started

Navigate to the Builder Lab by clicking Builder Lab in the sidebar. You'll see the Phase Stepper at the top showing your current progress.

You can start from a template to skip manual configuration. Click Start from Template and choose a pre-built agent configuration.

Phase 1: Discovery

The Discovery phase captures the fundamental definition of your agent.

Agent Metadata

Configure the basic agent information:

  • Name — A descriptive name for your agent (max 100 characters)
  • Description — What your agent does and who it's for
  • Category — The primary category (Code Assistant, Data Analysis, etc.)
  • Tags — Searchable keywords for marketplace discovery
  • Access Level — Public (anyone can use) or Private (only you)

Knowledge Base

Define the information your agent needs to operate:

  • Context Documents — Upload or paste reference materials
  • System Prompt — Instructions that guide the agent's behavior
  • Examples — Sample input/output pairs for few-shot learning

Persona Configuration

Optionally configure a persona for your agent:

  • Persona Name — A character name for the agent
  • Behavior Style — Formal, casual, technical, creative
  • Response Format — Markdown, plain text, JSON, etc.
  • Guardrails — Topics or actions the agent should avoid

Phase 2: Logic

The Logic phase uses a visual DAG (Directed Acyclic Graph) editor to define your agent's processing pipeline.

The DAG Editor

The DAG editor provides a canvas where you can:

  • Add Nodes — Click the canvas to add processing steps
  • Connect Nodes — Drag from one node's output to another's input
  • Configure Nodes — Click a node to set its parameters
  • Validate — The editor checks for cycles, disconnected nodes, and missing configurations

Node Types

| Node Type | Purpose | Example | |-----------|---------|---------| | LLM Call | Send a prompt to an AI model | Generate a response | | Transform | Modify data format or structure | Parse JSON, filter fields | | Conditional | Branch logic based on conditions | Route based on intent | | Service | Call an external API or service | Fetch data, send email | | Gate | Control flow validation | Verify data quality |

Validation Rules

Before proceeding to Synthesis, the pipeline must pass validation:

  • No circular dependencies (cycles)
  • All nodes reachable from the start node
  • Required inputs connected for every node
  • At least one output node defined

If validation fails, the editor highlights the problematic nodes and edges in red. Hover over the error markers for detailed explanations.

Phase 3: Synthesis

The Synthesis phase generates the final agent blueprint from your Discovery and Logic configurations.

Blueprint Review

Review the synthesized configuration:

  • Agent Configuration — Merged metadata, knowledge base, and persona
  • Pipeline Definition — Serialized DAG with node configurations
  • Resource Requirements — Estimated compute resources needed
  • Sample I/O — Preview with sample input and expected output

Publishing

When you're satisfied with the blueprint:

  1. Click Publish to open the publishing dialog
  2. Choose the initial version number (defaults to v1.0.0)
  3. Add release notes describing the agent
  4. Select marketplace visibility (Public or Private)
  5. Optionally set execution pricing

Published agents appear in the marketplace within seconds. Users can discover, star, and execute your agent immediately.

Auto-Save and Recovery

The Builder Lab automatically saves your progress:

  • Auto-save triggers every 30 seconds during active editing
  • Session Recovery restores your last state if you close the browser
  • Version History lets you revert to any previous save point

Advanced Features

BMAD Expansion Pack

For advanced users, the Builder Lab supports BMAD Expansion Pack generation:

  • Generate complete BMAD workflow packs from agent configurations
  • Export packs as downloadable ZIP files
  • Import packs from the marketplace to reverse-map into the canvas

Split-Screen Terminal

The Output Terminal in the Builder Lab provides:

  • Real-time synthesis output
  • Error logs and debugging information
  • Pack generation progress
  • Export controls

Next Steps