What is LLM (Large Language Model)?
The underlying neural network an agent uses to reason and generate text — e.g. Claude, GPT-4o, Qwen. The agent framework (OpenClaw, Hermes, etc.) is the scaffolding around an LLM that gives it tools, memory, and goals.
How it is built
A large language model is an AI language model with a very large number of parameters, trained on vast amounts of text. Anthropic describes the first stage, pretraining, as training on a large unlabeled corpus to predict the next word given the text so far. A pretrained model on its own is not inherently good at answering questions or following instructions and often needs careful prompting to behave.
How it becomes an assistant
Fine-tuning and reinforcement learning from human feedback (RLHF) refine a pretrained model so it follows instructions and acts more like a chatbot. In RLHF, humans rank example outputs and the model is rewarded for producing text like the higher-ranked ones. Claude, for example, is based on a large language model that has been fine-tuned and trained with RLHF.
What that means for agents
The model only produces text, or structured tool calls, from what is in its context window. Everything else an agent does, such as running commands, reading files or remembering things, comes from the software around the model. Even with temperature set to 0, outputs are not fully deterministic, so identical inputs can give different results across calls.
Example
You ask a model to summarise a changelog. It was pretrained to predict text, fine-tuned and RLHF-trained to follow the instruction, and it answers only from the changelog text you put in its context window, plus whatever it learned in training.
Sources
See also
← Back to the full AI agent glossary.