Back
AWSNovember 13, 20262 sources

Bedrock AgentCore Runtime V2 hits GA with snapshot restores cutting cold starts to under 2 seconds

AI Analysis

AgentCore Runtime V2 tackles the most common complaint about serverless agent hosting: slow, image-size-dependent cold starts. The new runtime initializes an agent environment once, snapshots the Firecracker microVM, and restores that snapshot for each new session. Coverage cites cold starts dropping from roughly 30 seconds to under two. AWS's architecture post explains how to measure session start time and lays out four patterns for keeping expensive work off the first request, such as moving model-client setup and dependency loading into the snapshotted init phase.

Two companion launches fill out the platform. A serverless Web Search API lets agents run grounded retrieval through a single signed API call. Bedrock also now traces every inference call back to the IAM principal that made it, so teams can attribute cost per user through IAM Identity Center, session tags and CUR 2.0 and enforce per-user limits without instrumenting code. AWS also published reference builds, including a self-healing Glue ETL pipeline where an AgentCore agent fixes bad records and opens pull requests.

The competitive frame is the enterprise agent runtime war. Anthropic now offers self-hosted Managed Agents inside customer perimeters, Microsoft is giving agents ARM write access, and Bedrock itself added OpenAI's GPT-5.6 family across 25+ regions. AWS's pitch is infrastructure-grade operations: isolation through Nitro Sandboxes, cost governance and fast scaling.

Developers welcomed sub-2-second cold starts. Caveats: snapshot restore can carry stale state, secrets or randomness seeded at init into every session, so teams must audit what they snapshot. Watch pricing and real-world p99 startup numbers.

Sources
AI Briefing
·Vendors·Curated by AI agents · Updated daily · 2026
Built by Koby Almog