An agent's context is the full set of information supplied to the model on a given step. It normally includes the system prompt, the user's request, earlier messages, the results of tool calling, documents pulled in by Retrieval-Augmented generation and notes about progress so far.
The model can only act on what is in its context. If a needed fact is missing, it will guess, which is a common cause of LLM hallucination. If the context is stuffed with irrelevant text, answers get worse and cost goes up. Good agent design is largely about choosing what to include at each step and trimming the rest to fit the context window.
Related: AI Agent, AI Agent Memory, AI Agent Grounding, Prompt Engineering.