Mastering the T20 Multi-Agent System

Chapter 5: The T20 Team: Agents in Detail

This chapter provides an in-depth look at each agent comprising the T20 Multi-Agent System. Understanding the specific roles, responsibilities, and underlying technologies of these agents is key to leveraging the system's full potential. We will also cover how agents are defined and how you can integrate your own custom agents.

5.1 Meta-AI (Orchestrator)

• Role: Orchestrator
• Core Responsibilities: Analyzes high-level goals, generates dynamic execution plans (JSON), delegates tasks to specialized agents, and manages overall workflow execution.
• Underlying Model: gemini-2.5-pro (chosen for its advanced reasoning and planning capabilities).

5.2 Lyra (Prompt Engineer)

• Role: Prompt Engineer
• Core Responsibilities: Optimizes agent performance by dynamically refining system prompts during workflow execution. Structures workflows and ensures clarity in agent instructions.
• Underlying Model: gemini-2.5-flash (selected for its speed and efficiency in prompt manipulation).

5.3 Aurora (Designer)

• Role: Designer
• Core Responsibilities: Generates aesthetic layouts, color palettes, typography suggestions, and UI/UX flow designs based on task requirements.
• Underlying Model: gemini-2.5-flash-lite-preview-06-17 (suitable for creative and visual output generation).

5.4 Kodax (Engineer)

• Role: Engineer
• Core Responsibilities: Implements designs into clean, modular, and performant code (e.g., HTML, CSS, JavaScript). Translates design specifications into functional implementation.
• Underlying Model: gemini-2.5-flash-lite-preview-06-17 (chosen for efficient code generation).

5.5 TASe (Task-Agnostic Step Extractor)

• Role: Task-Agnostic Step Extractor
• Core Responsibilities: Identifies and extracts fundamental, reusable 'Task Agnostic Steps' (TAS) that contribute to achieving high-level goals.
• Underlying Model: gemini-2.5-flash-lite-preview-06-17 (efficient for analytical tasks).

5.6 Defining Your Own Agents

The T20 system's flexibility is greatly enhanced by its declarative agent definition mechanism. Agents are configured using simple YAML files, typically located in the agents/ directory. This makes it easy to understand, manage, and extend the agent team.

Structure of Agent YAML Files

A typical agent definition file includes the following key fields:

• name: A unique identifier for the agent (e.g., Meta-AI).
• role: The primary function or expertise of the agent (e.g., Orchestrator, Designer).
• goal: A concise description of the agent's main objective.
• model: The specific Large Language Model (LLM) powering the agent (e.g., gemini-2.5-pro). This choice dictates the agent's capabilities.
• prompt (Optional): A default system prompt can be specified here, though Lyra may dynamically adjust it.

Integrating Custom Agents

To integrate a new custom agent:

1. Create a new YAML file in the agents/ directory defining your agent's properties (name, role, goal, model).
2. Ensure the underlying model is accessible and configured correctly.
3. The T20 system's loader will automatically recognize and incorporate your new agent into the available team, making it available for delegation by the Orchestrator.

This modular approach allows for significant customization and extension of the T20 system's capabilities.