The Context is the input that is passed to the Model for probabilistic completion. This is passed in as a sequence of tokens and then tokens are generated based on a probabilistic distribution as an output.
Order matters since preceding tokens in the context influence the tokens that are produced.
Each Model has a fixed Context Window size for both input and output. Anything falling outsize the maximum allowed tokens is either ignored or dropped.
It is important to remember that an LLM has no memory. It is stateless. So everything that is needed for the next projection is contained in the context. This includes system instructions, tool and MCP definitions, any included context from files or a repository, and of course the user Prompt.
AI Agents often use a technique called Context Compaction to manage a the size of the context. Modern Model context window sizes can be over a million tokens but their effectiveness drops off far sooner than that. This is often called Context Rot and is a core consideration in Context Management.