Mastering the T20 Multi-Agent System

Chapter 2: Core Concepts and Architecture

2.1 The Orchestrator-Delegate Model

The T20 system operates on a sophisticated Orchestrator-Delegate model. This architecture is central to its dynamic planning and execution capabilities. The Orchestrator agent, primarily Meta-AI, serves as the brain of the operation. Its core responsibilities include:

• Goal Analysis: Interpreting the high-level user goal.
• Plan Generation: Dynamically creating a step-by-step plan, often in JSON format, outlining the tasks required to achieve the goal.
• Task Delegation: Assigning specific sub-tasks to the most appropriate specialized agents (delegates) based on their defined roles and capabilities.
• Context Management: Ensuring that the output and context from one step are correctly passed to the next.

The Delegate Agents are specialized entities, each designed to perform a specific set of tasks. Examples include Aurora for design, Kodax for engineering, and TASe for step extraction. They receive tasks from the Orchestrator, execute them, and return their outputs (artifacts) back to the Orchestrator or the next agent in the sequence.

2.2 Dynamic Planning Mechanism

A cornerstone of the T20 system is its Dynamic Planning Mechanism. Unlike rigid, pre-programmed workflows, T20 generates plans on-the-fly for each unique goal. This process typically involves:

1. Orchestrator Interpretation: The Meta-AI agent analyzes the user's high-level goal.
2. LLM-Powered Planning: The Orchestrator queries a powerful LLM (like Gemini 2.5 Pro) to break down the goal into a sequence of actionable steps.
3. Structured Plan Output: The generated plan is structured, often in JSON format, detailing the sequence of tasks, the agents responsible for each task, and any dependencies.
4. Schema Enforcement: The system utilizes Pydantic for schema enforcement, ensuring that the plans generated by the LLM are reliable, machine-readable, and adhere to a predefined structure. This prevents errors and ensures smooth execution.

This dynamic approach allows T20 to adapt to a wide variety of tasks without requiring manual workflow configuration for each new objective.

2.3 Contextual Collaboration and Artifact Passing

Effective collaboration between agents is facilitated through Contextual Collaboration and Artifact Passing. As the Orchestrator executes the plan, the output of each agent's task is treated as an artifact. These artifacts represent the tangible results of an agent's work (e.g., generated code, design descriptions, extracted information).

• Stateful Sessions: Each run of the T20 system is contained within a dedicated session directory. This directory stores all artifacts, prompts, and logs generated during that specific execution.
• Context Chaining: Artifacts produced by one agent are made available as inputs for subsequent agents. This creates a chain of context, allowing agents to build upon each other's work seamlessly. For example, design specifications generated by Aurora can be passed to Kodax for implementation.

This mechanism ensures that information flows correctly through the system, maintaining a coherent state and enabling complex, multi-step task completion.

2.4 Meta-Learning and Prompt Engineering (Lyra)

The T20 system includes a unique agent, Lyra, the Prompt Engineer. Lyra's role is to enhance the performance and effectiveness of other agents by refining their system prompts *during* the workflow execution. This Meta-Learning capability allows the system to:

• Adapt Prompts: Adjust prompts based on the specific task, intermediate results, or desired output quality.
• Optimize Performance: Improve the accuracy, creativity, or efficiency of agent responses.
• Ensure Consistency: Maintain a consistent tone and adherence to instructions across different agents and tasks.

By dynamically optimizing prompts, Lyra contributes significantly to the overall robustness and adaptability of the T20 system.

2.5 Session Logging and Traceability

Session Logging is a fundamental feature of the T20 system, providing unparalleled Traceability. Every aspect of an execution run is meticulously recorded within the session directory:

• Initial Goal: The original user input.
• Initial Plan: The AI-generated plan (initial_plan.json).
• Agent Prompts: The exact prompts sent to each agent.
• Agent Outputs: The artifacts (results) generated by each agent.
• Execution Logs: Detailed logs of agent actions and system operations.

The structure of these session directories (e.g., sessions/session_<uuid>/) makes it easy to locate and review specific runs. This detailed logging is crucial for:

• Debugging: Identifying the root cause of errors or unexpected behavior.
• Auditing: Understanding the complete decision-making process.
• Learning: Analyzing how the system tackled specific problems to improve future interactions.

This commitment to transparency and traceability ensures that users can fully understand and trust the T20 system's operations.

Key Features Summary:

• Declarative Agent Definition: Agents configured via simple YAML files.
• Dynamic, AI-Generated Plans: Custom plans created for each goal.
• Role-Based Delegation: Tasks assigned to the most suitable agents.
• Structured LLM Outputs: Pydantic used for reliable, machine-readable plans.
• Stateful Sessions: Isolated directories for each run's artifacts and logs.
• CLI-Driven: Easy operation via Command-Line Interface.
• Powered by Google Gemini: Leverages Gemini models for orchestration and specialized tasks.