AWS Launches Reworked Bedrock AgentCore Runtime for Elastic Memory

Amazon Web Services announced a reworked version of the managed compute layer in Amazon Bedrock AgentCore on September 18, 2026. [2] The new AgentCore runtime reclaims memory as agent sessions release it and delivers consistent cold start times regardless of container image size or concurrency. [2] The platform scales down to zero when an agent has no work, and billing follows resource utilization with no charge for idle CPU waiting on I/O. [2]

Under the original runtime, a session held its allocated memory from allocation until the session ended. [2] This left long-running or bursty agents paying for their high-water mark well after the memory stopped being used. [2] Cold start latency previously grew with image size and concurrency, peaking under bursty traffic when the fewest ready environments remain. [2] Now, a session that lands on an already-initialized environment begins in under 100 milliseconds. [2]

By introducing elastic memory reclamation and consistent cold starts to Bedrock AgentCore, AWS addresses critical cost and latency bottlenecks for bursty, long-running agent deployments. The shift away from paying for peak memory marks a maturation of serverless economics for generative AI, aligning compute costs closer to actual usage and mitigating the latency penalty of scaling up during traffic bursts. What the sources don't address: How much the elastic memory reclamation will reduce average compute costs for typical enterprise agent workloads compared to the legacy version.