New Bedrock AgentCore Runtime GA with elastic memory and fast cold starts

AWS announced general availability of the new AgentCore Runtime, the serverless microVM compute layer inside Amazon Bedrock AgentCore. The redesign targets two problems that made always-on agents expensive: idle agents previously held peak memory for the whole session, and cold-start latency varied unpredictably with image size and concurrency. The new runtime reclaims unused memory throughout a session — billing only for active memory — and uses snapshot-based starts to deliver consistent cold-start times at any scale.

VP of Agentic AI Swami Sivasubramanian framed the rationale on the shift to ambient, always-on agents: 'More and more agents now process claims, write and review code, coordinate across systems, and run for hours without supervision. As agents become ambient and always-on, the infrastructure underneath them has to be just as capable.' The runtime is fully serverless, with consistent fast session starts 'whether you're running one agent or thousands.'

The launch is part of a broader AWS agent-infrastructure push unveiled around its Global Meeting — including migration guides moving multi-model agents off self-managed ECS/Fargate onto AgentCore. Competitively, it's AWS answering the operational-cost objection to agentic deployments head-on, staking Bedrock as the managed home for production agents against self-hosted Kubernetes stacks and rival clouds. The 'pay for active memory' model is the sharpest hook — it directly attacks the economics that made long-running agents cost-prohibitive. Enterprises will want to see real bills before believing the savings, but the positioning is clear: AWS wants agents to be as boring and cheap to run as Lambda functions.