AWS Explains Bedrock AgentCore V2 Snapshot Startup: Sessions Ready in ~2 Seconds
AWS ML Blog · rss · 2026-11-14
How AgentCore runtime V2 speeds up session startup
Amazon Bedrock AgentCore runtime gives agents a fully managed serverless environment: each session runs in an isolated microVM, scales to zero, and bills only for consumed resources. The new platformVersion V2 changes session startup: instead of initializing the container per session, it prepares the environment once, snapshots it, and restores that snapshot for each new session — faster, more consistent starts, plus memory reclamation so bills track actual usage.
Key facts about the snapshot mechanism
- At build time: container starts, waits for the first healthy ping (within 120s), file cache is cleared, and the microVM is frozen with memory and disk saved as a snapshot; create/update now takes a few minutes.
- On restore, the process resumes exactly where frozen — imports, objects, and clients like boto3 are in memory with endpoint and credential resolution done. File cache does not carry over; image files are fetched on first read.
- Network connections opened before the snapshot go stale — close them.
Four patterns by scenario
- 2s is fine: switch to V2, do all init before app.run().
- Need large files/CLI/browser that can't live in-process: use the new build hooks to prime them during snapshot build.
- Need sub-second and users open the UI first: pre-warm sessions in the background on user engagement.
Measure before optimizing
- Split Latency (TTFB) vs OverheadLatency (platform) in CloudWatch; the difference is your code's time.
- Report P90/P99 or trimmed means like TM90; test at production new-session rates.
- On V2, init runs once at snapshot build, not per request — log it separately at startup.
The approach mirrors AWS Lambda SnapStart and is especially useful for latency-sensitive chat/voice agents.
More from coding & agent
- ClaudeAI weekly: Fable 5.2 burns tokens, secrets leak into repos, usage eats failures — ClaudeAI-mod-bot · 2026-11-14
- 97% of coding-agent tokens are cache reads, but cache writes eat 31% of the cost — Nitjsefnie · 2026-11-14
- 4B specialized model hits 93.8% on bank support, beating proprietary models' 74% — MParakhin · 2026-11-14
- EnvOpt: Optimizing Agent Environments Beats Bigger Models on Cost and Pass Rate — AlexGDimakis · 2026-11-14
- workerd lands KV, Queues and Workflows in cluster, first release approaching — irvinebroque · 2026-11-14
- Researcher uses a coding agent to fetch latest Marin scaling-law fit parameters — RylanSchaeffer · 2026-11-14