Modern AI agents are not powered by a single large language model. Production-grade agents built
using an Agent Development Kit (ADK) rely on multiple specialized model types, each designed for
reasoning, planning, action execution, perception, and memory. This document provides a clear
overview of the key model categories commonly used in ADK-based agent architectures.
GPT – Generative Pre-trained Transformer
GPT models serve as the foundational language intelligence of an agent. They are responsible for
understanding user intent, handling natural language interactions, and generating coherent responses.
In ADK systems, GPT models often act as planners or high-level reasoning components.
MoE – Mixture of Experts
Mixture of Experts architectures dynamically route tasks to specialized sub-models. This approach
enables better scalability, lower inference costs, and improved performance by selecting the most
relevant expert for each task.
LRM – Large Reasoning Model
Large Reasoning Models are optimized for multi-step logical reasoning and complex decision-making.
They are particularly effective in scenarios that require planning, evaluation of alternatives, and
long-horizon problem solving.
VLM – Vision Language Model
Vision Language Models extend an agent’s capabilities beyond text by enabling visual understanding.
They allow agents to interpret images, documents, diagrams, screenshots, and other visual inputs
within a unified framework.
SLM – Small Language Model
Small Language Models are lightweight and efficient, making them suitable for high-frequency or
low-latency tasks. They are commonly used for simple classifications, background processing, and
deployment on resource-constrained systems.
LAM – Large Action Model
Large Action Models enable agents to interact with external systems. They are responsible for invoking
APIs, using tools, executing workflows, and performing real-world actions based on model outputs.
HLM – Hierarchical Language Model
Hierarchical Language Models structure agent behavior into multiple layers, typically following a
manager–worker paradigm. A high-level planner decomposes tasks, while specialized worker agents
execute individual subtasks in parallel or sequence.
LCM – Large Concept Model
Large Concept Models focus on semantic understanding at the concept level rather than surface-level
text. They support long-term memory, contextual grounding, and knowledge abstraction, enabling more
consistent and human-like reasoning.
Key Takeaways
Effective AI agents are built by combining multiple specialized models, each with a clearly defined role.
The Agent Development Kit is not merely a collection of tools, but an architectural approach to
designing scalable, robust, and production-ready agent systems.
