The context window is the complete set of information a language model processes to generate responses. It includes all your messages, the model's replies, tool calls, and even internal reasoning outputs. As conversations grow, so does the context window, which can impact the model's performance. Every new message you send is added to this window, and when the model generates a response, it analyzes everything in that window, not just your latest input. Models have limits on how much context they can handle. If a conversation exceeds this limit, the model can't process any more input. Furthermore, all text in the context window transforms into tokens for processing, meaning every piece of text influences the output. Too much context can degrade quality; shorter, focused conversations typically yield better results. As conversations lengthen, the chances of the model producing inaccurate or irrelevant outputs increase. In the Amp environment, the context window combines your conversation with an agent, which consists of a model, system prompts, and tools. The system prompt dictates how the model interacts and uses tools, while environmental data, like the operating system or open files, also feeds into the context window. This makes the management of context crucial since altering any component can significantly change the agent's behavior. Amp offers several ways to manage the context window. You can add context by mentioning files, using Shell Mode for terminal commands, or asking the agent to use specific tools. Editing previous messages or restoring the conversation to an earlier state allows you to refine the context. Handoff features let you distill information from one context window to another, ensuring relevant details are preserved while irrelevant ones are discarded. This management capability is key to optimizing interactions with the agent and ensuring effective and accurate outputs.