Glossary
Context compaction
A harness step that replaces older turns of an agent transcript with a shorter summary so later model calls still fit.
Context compaction is the harness step that shortens a long transcript. Later calls see the summary, plus whatever recent turns the harness keeps, instead of the whole history.
Two product versions were re-read on November 9, 2026. Anthropic’s context-window page calls server-side compaction the primary strategy for long-running agent work. The threshold page says compaction starts when input tokens reach a trigger, 150,000 by default and at least 50,000, and that threshold compaction is in beta on select models. OpenAI’s compaction guide says the carried item is opaque and not intended to be human-interpretable.
A research handoff is a third version. The November 3, 2026 report You don’t need long context has the model write a short brief and then starts a new conversation. The space being protected is the context window. The text of one call is a prompt. The cycle that fills the window is the agent loop.