Frequent context compaction shows little accuracy loss in Nvidia's AI-agent tests

Nvidia's researchers had agents write a short handoff note and restart with it and their task, rather than carry the full transcript. Across six long-horizon benchmarks, at most five of 22 compaction settings per model lost more than two points versus full history. The paper also reports that serving open models with a 256K context window cost 3.6 to seven times more per generated token than with 16K.