Sources checked:
What changed
AWS announced AgentCore Runtime V2 on 18 September, reporting elastic memory reclamation and snapshot-based starts. These are vendor claims, not independently measured savings for a customer workflow.
Business implications · analysis
Commercial interpretation: the useful comparison is a completed business task, not an isolated infrastructure statistic. For an agent that prepares a supplier comparison, count model calls, external tools, waiting time, retries and human corrections alongside runtime charges. Keep the same input set, acceptance criteria and permissions when comparing versions. A cheaper run that needs more checking can be a more expensive process. Test both a short request and a longer session with pauses; an average alone can hide the cases that consume most of the budget.
What to watch next
Watch cost per accepted result, slow-case completion time and failure recovery in a bounded pilot. If overhead falls without weakening accuracy, wider use may be justified. If integration or review dominates, improving those steps may matter more than changing the runtime. This is a testable operating scenario, not a promised productivity gain.
Primary sources
Company statements are attributed to the publisher. Our analysis is separate; forward-looking points are questions to monitor, not promised outcomes.