Agent Context Management

An agent can have access to far more information than should fit into one model request. Context management selects what matters now.

Separate stable and dynamic context

Stable instructions, tool definitions, recent conversation, retrieved documents, and tool results have different lifecycles. Keep them logically separate so each can be compressed or cached appropriately.

Summarize old history

Preserve decisions, commitments, IDs, and unresolved tasks rather than every token of old conversation.

Trim tool output

Do not feed megabytes of API payload back to the model. Return structured summaries plus identifiers for follow-up retrieval.