There is a cliff in agent system-prompt compression: safe to 75% retained, catastrophic below 35%, and the method you pick swings success from 47% to 20%
CompressAgent is an environment-verified benchmark for compressing agent control contexts — the persistent system-side instructions specifying tools, arguments, policies, execution protocols and recovery — across nine independently built contexts and 15,525 runs. At 75% retained context, generic rewriting and section-based compression hold 92.7% and 92.4% success against a 93.8% full-context baseline; between 50% and 35% the methods diverge sharply, and at 35% section-based scores 47.0% while generic rewriting collapses to 19.9%; from 25% down, executable protocols become fragile outright. Failures surface specifically as tool-execution and action-parsing errors, and reliability varies enough across contexts that universal compressor rankings are inappropriate — if you trim your agent's system prompt, qualify each compression per-context against executable outcomes, not token counts.
↳ Follow the thread