Amazon has updated Bedrock AgentCore Runtime to start isolated AI-agent sessions in under two seconds while supporting workloads up to 8 vCPUs, 32 GB of memory and 10 GB container images. AWS says that is four times the previous compute ceiling and as much as 15 times faster to start.
Every session runs in a separate Firecracker micro virtual machine with its own identity, access policy and audit trail. The new version can mount working data from storage such as an S3 bucket through FUSE, extending the session’s controls to files that agents create or modify during a long task.
Billing and resource use now follow the stop-and-start pattern of agents more closely. Memory can be reclaimed after an agent releases it, and an idle session can scale to zero. AWS claims this can reduce memory consumption by up to 68% for multi-turn workloads with pauses; customer examples in the announcement report lower cold-start time and memory use, but the results are vendor-selected rather than independent benchmarks.
The runtime is available across AWS regions in the Americas, Europe and Asia Pacific. Organizations still need to define tool permissions, data boundaries and approval rules. Hardware isolation separates sessions, but it does not determine whether an agent’s intended action is appropriate.