Context Package

A context package is the rich bundle of conversation state that an interaction model sends to a background reasoning model when delegating a task that requires deeper processing than the interaction model can produce in real time. Unlike a standalone query — which contains only the current user input — a context package includes the full conversation history, current state of all active modalities, accumulated context, and any other information the interaction model has gathered that might be relevant to the delegation. This ensures that the background model has everything it needs to reason deeply about the task without requiring the user to re-explain context that the interaction model already understands.

The context package is a key architectural innovation because it separates the responsibilities of real-time responsiveness and deep reasoning. The interaction model must respond within milliseconds to maintain the sense of presence and immediacy that makes continuous interaction feel natural. But some tasks — complex reasoning, multi-step tool use, extensive search — require more computation than can be completed within that latency budget. Rather than blocking the interaction to do deep reasoning inline, the interaction model packages what it knows and hands it off, continuing to remain present and responsive to the user while the background model works.

When the background model produces results, they do not simply replace the current context or appear as an abrupt interruption. Instead, the interaction model receives a stream of updates as the background model produces them, and integrates these into the ongoing conversation at a moment that is contextually appropriate — when the user has paused, when the current topic is still active, or when the user has signaled readiness for a substantive update. This contrasts with conventional agent systems, where a tool call or retrieval operation typically produces a discrete result that the model then incorporates into its next response, often after a noticeable delay and context switch.

The design of the context package — what to include, how to structure it, how to handle partial updates, and how to manage the state synchronization between interaction and background models — is an active area of research. The challenge is that context packages grow with session length, and managing memory and relevance over very long sessions is unsolved. The interaction model must also decide what to delegate versus what to handle immediately — a judgment that requires understanding both the task's complexity and the user's current focus.