LLM Agentic Playground
======================

This page introduces the LLM Agentic Playground, a modular environment to test and evaluate multi-agent setups for control of liquid cooling systems.

.. image:: ../images/explainable.PNG
   :scale: 60%
   :alt: Agentic LLM System Overview
   :align: center

Configurable Reasoning for Trade-Off Testing
--------------------------------------------

Users can configure different reasoning parameters to explore the trade-offs between decision quality, latency, and memory usage.

- **LLM Models:** Llama 8B, Qwen 8B
- **Context Window Sizes:** 8K to 128K
- **Reasoning Modes:** Chain of Thought, Few-Shot, Extended Thinking

Drag-and-Drop Agent Design for Scalable Control
-----------------------------------------------

A visual agent builder enables users to define scalable multi-agent architectures. These can be used for both **coarse-** and **fine-grained** control strategies in digital twins for liquid cooling.

.. image:: ../images/agentic-builder.PNG
   :scale: 40%
   :alt: Agentic LLM System Overview
   :align: center

Stress Injection for Reasoning Evaluation
-----------------------------------------

System stress can be simulated through fault injection, allowing users to evaluate how various agent setups and reasoning styles impact:

- System resilience
- Fault recovery time
- Control robustness

Real-Time Monitoring
--------------------

Dashboards provide visibility into:

- Agent and message flow
- Event alerts
- Key performance indicators (KPIs) related to liquid cooling
- Control latency and dynamics

Explainability Tools
--------------------

The system includes natural language explanation tools that provide:

- Short-form rationales
- Detailed reasoning traces

These help users understand decision-making processes and compare performance across configurations.

Safety and Guardrails
---------------------

Built-in safety constraints ensure that all tests remain within operational boundaries. This allows users to:

- Explore aggressive optimization strategies
- Avoid causing system damage