New AgentCore Runtime Goes GA on Amazon Bedrock with Elastic Memory, Fast Cold Starts
AWS's next-generation AgentCore Runtime targets the economics and reliability of always-on agents. Built on serverless microVMs, its headline features are elastic memory management — reclaiming unused memory during a session so customers pay for actual consumption rather than provisioned peak — and consistent cold-start times regardless of container image size or concurrency. Both address pain points that make production agents expensive and unpredictable.
The mechanism matters for the agentic shift AWS's Swami Sivasubramanian described: agents that 'process claims, write and review code, coordinate across systems, and run for hours without supervision.' Long-running, bursty memory profiles are exactly where pay-for-peak pricing punishes users, so per-session reclamation directly cuts cost. Consistent cold starts remove the tail-latency lottery that plagues serverless at scale.
Competitively, this is AWS fortifying Bedrock as the agent-infrastructure layer against Azure AI Foundry (which had a rough week with a CVSS 10.0 patch) and Google's Vertex, and it dovetails with AWS's parallel launches — Lambda MicroVM agent sandboxes, SageMaker HyperPod Inference Gateway, and Kimi K3 GA. The through-line is AWS positioning for the production-agent era that NVIDIA's Huang and xAI's Grok 4.8 also invoked.
Swami's LinkedIn announcement drew 1,213 likes, signaling strong developer interest, though the real test is production economics under load. Skeptics will want to see whether elastic reclamation introduces latency or instability mid-session, and whether 'consistent' cold starts hold at extreme concurrency. Watch for independent cost comparisons versus the prior runtime and versus self-managed ECS/Fargate setups the migration guides target.
