Large Language Model (LLM)

A large language model, or LLM, is a neural network trained to model patterns in sequences of tokens. Given an input, it predicts likely next tokens and can generate text, code, structured data, and other outputs.

How an LLM works at a high level

Text is divided into tokens, which may represent words, parts of words, punctuation, or other units. The model processes the input tokens within its context window and calculates probabilities for what should come next.

By generating one token after another, the model can answer questions, summarize documents, translate text, write code, classify information, and follow instructions. Its behavior comes from patterns learned during training plus the context supplied at inference time.

An LLM is not automatically an agent

An LLM normally produces an output for a given input. An AI agent adds a surrounding system that can maintain state, use tools, observe results, and repeat actions toward a goal.

This distinction is central to What Is an AI Agent?. The model may serve as the reasoning and language engine, but the agent also needs control logic, permissions, tool execution, and stopping conditions.

For example, an LLM can suggest a SQL query. A data agent can inspect a schema, generate the query, execute it through a tool, observe an error, revise the query, and return validated results.

LLM knowledge has limits

An LLM’s internal knowledge is learned from training data and is not a live database. It may be incomplete, outdated, or incorrect. The model can also generate plausible statements that are not supported by evidence.

Agent systems address some of these limits by retrieving current information, calling tools, validating outputs, and asking humans for approval. These controls reduce risk but do not make the model infallible.

Model capability is only one design choice

Different agent tasks need different balances of reasoning quality, speed, context capacity, cost, tool-calling reliability, privacy, and deployment control.

A more capable model can improve decisions, but reliable agent behavior also depends on clear instructions, well-designed tools, good state management, useful observations, and bounded autonomy. The surrounding architecture often matters as much as the model itself.