AgentCore Runtime V2 drops agent cold starts from 30s to ~2s
AWS's Sep 18 agent-infra push also bills memory by real use, not peak — and a new GPU-aware gateway cuts self-hosted first-token latency up to 98%.
Copy markdown
Cold starts fall to about 2 seconds
The new AgentCore Runtime (set `platformVersion` to V2) posts a P75 cold start of 1.9–2.0s for container images from 200MB to 2GB, versus 5.4–30s on V1. Scale-to-zero agents feel instant now, so you can stop paying to keep a warm instance idle just to dodge the lag.
Pay for the memory you use, not the peak
V2's elastic memory reclaims unused RAM mid-session and bills actual usage instead of the peak you reserved for the whole run. A session that spikes once no longer pays for that spike across its entire lifetime. Live in us-east-1/2, us-west-2, eu-west-1 and ap-northeast-1.
GPU-aware routing lands for self-hosters
SageMaker HyperPod's new Inference Gateway is a Kubernetes-native add-on that routes each request by live GPU signals — KV-cache use, queue depth, LoRA residency — instead of round-robin. AWS clocked up to 98% lower P95 first-token latency and +50% throughput on Qwen3-32B across mixed GPUs.
Elsewhere: Vercel trims the deploy-and-eval loop
Vercel's CLI now ships static artifacts in under a second, and Harbor 0.22.0 can run Terminal-Bench and other agent evals directly on Vercel Sandbox. Tighter loops for anyone iterating on agents or shipping preview builds.