Amazon Bedrock’s AgentCore runtime has been refreshed to deliver a faster, larger, and more efficient isolated agent runtime. Sessions now start in under two seconds, can provision up to 8 vCPUs and 32 GB of memory, and accept container images as large as 10 GB, all while preserving the per‑session Firecracker microVM isolation guarantee.
Performance and scaling improvements
The new runtime claims up to a 15× reduction in start‑up time compared with the previous version. Compute capacity per session has been increased fourfold, enabling workloads that require compilation, browser automation, or local model inference to run without additional provisioning steps. Reported efficiency gains of up to 68% mean that the same workload consumes fewer CPU cycles for the same amount of work.
Customer observations reinforce these numbers. One user measured a 35% faster cold start for inference requests and a 77% faster start for health‑check calls, while overall cost per call dropped by 17%. Another reported a roughly 60% reduction in memory footprint per invocation, allowing higher traffic volumes at steady latency.
Operational impact
Because sessions launch quickly regardless of image size, the need for pre‑warmed pools or dedicated clusters disappears. Teams can rely on pure on‑demand provisioning, which eliminates idle resource consumption and simplifies capacity planning. The ability to run larger container images (up to 10 GB) removes a previous barrier for agents that bundle extensive libraries or model binaries.
Practically, this translates to:
- Lower baseline cost: no idle VMs waiting for the next request.
- Predictable latency: sub‑2‑second start times even under high concurrency.
- Simplified CI/CD pipelines: build once, run anywhere within the same runtime limits.
Security posture
The core security model remains unchanged. Each agent session continues to run inside its own hardware‑isolated Firecracker microVM, with a distinct identity, policy set, and audit trail. This isolation boundary is still the primary defense against cross‑session data leakage or code execution interference.
Implications for security engineers include the need to maintain per‑session policy definitions and to ensure audit logging is integrated with existing monitoring stacks. The runtime does not alter the underlying isolation mechanism, so existing threat models around microVM escape remain applicable.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Teams should reassess their current agent deployment strategy. If you are using pre‑provisioned sandboxes, consider migrating to the updated AgentCore runtime to eliminate idle capacity and reduce cold‑start latency. Update capacity forecasts to account for the higher per‑session compute limits and the ability to run larger container images. Verify that your policy automation pipelines can generate the required identity and policy objects for each session, and confirm that audit logs are being collected at the microVM level.
Key actions:
- Benchmark your existing agent workloads against the sub‑2‑second start target.
- Adjust CI/CD to produce container images ≤10 GB and test them in the new runtime.
- Review and, if needed, expand per‑session policy definitions to cover the larger resource envelope.
- Integrate microVM audit streams into your security information and event management (SIEM) solution.
By aligning architecture and operational processes with these runtime changes, engineers can achieve faster, more cost‑effective, and securely isolated AI agent deployments at scale.