Manage agents by token count, not context-window length — 'context rot' begins ~32k tokens and tokens explain 80% of performance variance

Digital Applied·high signal

Databricks observed correctness degrading around 32k tokens even for large-window models, so a 200k window is not 200k of usable reasoning; Anthropic separately found token usage explains ~80% of multi-agent performance variance — more than model choice or tool-call count. The actionable inversion: optimize context management before reaching for a bigger model or window. Track tokens-in-play as your primary health metric and compact aggressively past the rot threshold.