New AgentCore Runtime in Amazon Bedrock delivers elastic memory and fast cold starts

AWS shipped a redesigned AgentCore Runtime, the serverless microVM compute layer inside Amazon Bedrock AgentCore, now generally available. Its two headline properties: elastic memory that reclaims resources as an agent session's needs shrink — so billing tracks actively used memory rather than peak allocation — and consistent cold-start latency regardless of container image size or concurrency, whether running one agent or thousands.

The design targets a specific shift AWS VP Swami Sivasubramanian described in an accompanying post (195 likes on X, 1,100+ on LinkedIn): agents that 'process claims, write and review code, coordinate across systems, and run for hours without supervision.' As agents become 'ambient and always-on,' the infrastructure economics of peak-memory billing and unpredictable cold starts become a real cost and latency problem. Elastic reclamation directly attacks the cost side by not charging for memory a session has stopped using.

The launch was part of a dense AWS AI week that also included Kimi K3 arriving on Bedrock, the SageMaker HyperPod Inference Gateway, open-weight models in the AWS European Sovereign Cloud, and Lambda MicroVM agent sandboxes — a coordinated push to own the agent-infrastructure layer rather than just model hosting. Matt Wood, AWS's Chief AI and Technology Officer, framed the strategy at the AWS Global Meeting: 'What we are doing today at AWS for AI is exactly what we did for cloud computing,' naming Trainium, Bedrock, and AgentCore as the building blocks.

Competitively, elastic per-session memory billing is a differentiator versus fixed-allocation serverless runtimes, and consistent cold starts address a common complaint in agentic deployments. The caveat: real-world savings depend heavily on session memory profiles, and the 'pay for actual usage' benefit is largest for spiky, long-running agents rather than short synchronous calls.