Next-gen AgentCore runtime becomes Bedrock default: sub-2s session starts, 8 vCPU/32GB microVMs

AWS has made its rebuilt AgentCore runtime the default for new Bedrock agent deployments. AWS VP of Agentic AI Swami Sivasubramanian summarized the gains on X: sessions start up to 15x faster, scale to 4x more compute, and run up to 68% more efficiently. Per AWS's What's New post:

- P75 session start latency dropped from 29.7 seconds to under 2 seconds.

- Sessions now support up to 8 vCPUs and 32 GB of memory.

- Container images can be up to 10 GB.

The mechanism is snapshot restore. Instead of cold-booting a container, AgentCore resumes a session from a snapshot inside a Firecracker microVM, the same isolation technology behind Lambda. Each agent gets a hardware-enforced boundary, which matters when agents run model-generated code. Idle sessions scale down so customers stop paying for memory they are not using. Third-party coverage says the rollout spans 19 regions. It arrived alongside related agent infrastructure: Cognito OAuth token exchange for agent delegation, Nitro Sandboxes for EKS, ECS and EC2 announced by Matt Wood, and the new AWS Emulator for local agent testing.

In competitive terms, long-running, larger agents are where Microsoft (Copilot Studio and Foundry), Google (Vertex agent engine) and specialist sandbox vendors are all heading. AWS is betting that isolation plus fast cold starts becomes the default way to host autonomous agents. Customer stories published the same day back that up: Natera's SWARM lab-ops agent on AgentCore reports roughly 130x faster first responses and 96.2% ticket coverage.

The caveats are practical. Existing runtimes are not described as auto-migrated. The 68% efficiency figure depends on idle patterns. Real-world P99 latency was not disclosed. Teams should also benchmark how state-heavy agents behave on snapshot restore, and verify that cost drops actually show up in CUR reports.