Microsoft and Microsoft Research introduced environment-probing curation, which gives agent memory curators read-only access to live environments, nearly doubling Copilot benchmark task success and halving inference costs.