Amazon has updated the managed runtime behind Bedrock AgentCore to reduce two costs of long-running AI agents: unpredictable cold starts and memory that remains billed after an application stops using it. The runtime now begins with a smaller memory footprint, adds pages on demand and reclaims cold or released memory during the session rather than retaining its peak allocation.
For startup, AgentCore initializes a container once, waits for it to become healthy and saves a snapshot. New instances restore that prepared environment instead of repeating the full boot and initialization sequence. AWS says this keeps starts consistent even as container images grow or many sessions arrive together.
In an AWS test of an empty echo agent, the new runtime delivered a 75th-percentile cold start of about two seconds for images from 200MB to 2GB. The previous version ranged from roughly 5.4 seconds to nearly 30 seconds as images became larger. The test used 5,000 cold invocations per configuration across two AWS regions, so it measured network travel as well as platform startup; it did not include model calls or real agent work. The update should matter most for interactive agents and bursty, long-lived jobs, but application-level latency will still depend heavily on the model, tools and workflow layered above the runtime.