Understanding Agents in Generative AI: A Class and Instance Definition

In the world of Generative AI and Large Language Models (LLMs), the term "agent" is becoming increasingly important. As of today, July 28, 2026, two key standards are shaping how we understand and implement agents in AI systems. This article will break down what an agent is, how it functions, and why this matters in the current tech landscape.

What is an Agent?

An agent in the context of AI is a system that can autonomously perform tasks by using tools and interacting with its environment. However, defining what exactly constitutes an agent is not straightforward. Two major protocols have emerged to guide this understanding:

1. A2A (Agent-to-Agent) Protocol v1.0: This standard, developed by the Linux Foundation, focuses on how agents communicate and integrate with each other. It specifies the interface of an agent but does not define what an agent actually is.
2. MCP (Model Context Protocol): Finalized today, MCP outlines how an agent uses tools. It details the internal workings of an agent but, like A2A, does not define the agent itself.

Both protocols highlight a gap in our understanding: neither clearly defines what an agent is. This article aims to fill that gap by providing a structured definition.

The Framework for Defining Agents

To create a comprehensive definition, we can think of agents in two parts: the class and the instance. The class represents the definition of an agent, while the instance is a specific realization of that class.

Agent Definition (Class)

The class of an agent includes several key components:

• Host Loop: This is the core process where the LLM (Large Language Model) decides actions, observes results, and iterates on its tasks. Without this loop, you simply have a chatbot that can use tools, not a true agent.
• System Context: This includes static elements such as persona and instructions that guide the agent's behavior.
• MCP Server Set: This defines which MCP servers the agent can interact with.
• Capability Boundary/Tool Policy: This outlines the conditions under which the agent can use certain servers and tools.
• LLM: The specific model that the agent employs.
• Termination Limits: These are constraints like maximum iterations and token budgets that govern the agent's operations.
• Task Contract: A schema that details what inputs the agent accepts and what outputs it must return.
• Context-Size Strategy: This describes how the agent manages its context size, including methods like compaction or top-x selection.

Agent Instance (Object)

The instance of an agent is defined by:

• Goal: The specific task or query initiated by the user.
• Working Context Window: This includes the conversation history and results from tools relevant to the current task.
• Credentials: Authentication details, such as OAuth tokens, that represent the identity of the user.
• Consumed Budget: Metrics that track the elapsed iterations and tokens used during the task.

The 2026 MCP Specification

The new MCP specification introduces significant changes, particularly regarding how state is managed. Previously, servers would remember sessions, but now, each call is stateless. This means that all necessary information travels with explicit handles that the model can see.

This change impacts how we define the working context window. The instance state now includes the goal, context, handles, credentials, and consumed budget. As a result, an agent can be treated as a serializable value, allowing it to be suspended, persisted, and resumed in different processes.

Conclusion

Understanding what an agent is and how it operates is crucial in the rapidly evolving field of AI. By distinguishing between the class and instance of an agent, developers can create more effective and adaptable AI systems.

Merits

• Clear framework for defining agents in AI.
• Facilitates interoperability between different AI systems.
• Supports the development of robust agent architectures.

Demerits

• Current definitions may not cover all real-world scenarios.
• Implementation may vary across different platforms and use cases.

Caution

This article is meant for educational purposes. Any placeholder values mentioned should be replaced with actual data in practice. Readers should verify claims against the original source before relying on them.

Frequently asked questions

• What is an agent in AI? — An agent is a system that can autonomously perform tasks by using tools and interacting with its environment.
• What are the A2A and MCP protocols? — A2A defines how agents communicate, while MCP outlines how they use tools, but neither defines what an agent is.
• What is the host loop in an agent? — The host loop is the core process where the LLM decides actions and iterates on tasks.
• What is the significance of the 2026 MCP specification? — It introduces a stateless model for managing agent interactions, impacting how agents maintain context.
• How do agents manage context size? — Agents use strategies like compaction or top-x selection to manage their context size effectively.
• Can agents be serialized? — Yes, the new MCP spec allows agents to be treated as serializable values, enabling easier persistence and resumption.

Tags

#agents #ai #mcp