Compaction: distill agent context windows before 100% to cut tokens 84% on long-running tasks

Anthropic Engineering·high signal

Anthropic's compaction technique high-fidelity-summarizes a context window when it nears its ceiling, letting an agent continue with minimal performance loss; a 100-round web search dropped token consumption ~84%. Budget compaction to trigger before 100% capacity rather than after, and pair it with structured note-taking so the agent reloads only distilled state.