Bernstein 2.16 brings cost attribution, safer summaries, and fleet control for multi-agent runs

The 2.16 release of the Bernstein orchestration framework targets operators running large fleets of AI agents, addressing issues that only emerge at scale. A new response-style profile system lets operators assign verbosity levels — terse, standard, or expansive — to individual roles, preventing over-spending on output tokens for tasks that need only brief results. Cost reporting now groups spending by profile, so operators can see exactly which role types are driving bills rather than receiving a single unexplained total. A built-in A/B harness allows side-by-side comparison of model arms by both cost and quality gate results, ensuring cheaper options are evaluated against measurable outcomes before adoption. The update also introduces improvements to summary reliability and lifecycle controls, giving operators greater ability to audit, pause, and reproduce multi-agent runs.