AWS ships AgentCore Runtime V2, cutting AI agent cold starts by up to 15x

AWS has shipped a rebuilt version of AgentCore Runtime, the serverless compute layer inside Amazon Bedrock AgentCore that hosts AI agents, cutting cold-start times by roughly 3x to 15x and moving to pay-for-actual-usage memory billing.

What's new

In a September 18, 2026 post to its "What's New" feed, AWS announced general availability of the next generation of AgentCore Runtime. The headline change is elastic memory management: the runtime now allocates memory dynamically during a session and reclaims what an agent isn't using, rather than billing for a fixed peak allocation, so customers "pay for actual usage rather than the peak."

The second major change is startup performance. Using snapshot technology, AWS says the new runtime "prepares the agent environment once and snapshots it" so that cold starts stay consistent regardless of how large or complex the agent's container image is. AWS's own benchmark numbers show the scale of the improvement: "P75 cold start of 1.9 to 2.0 seconds for container images from 200 MB to 2 GB, compared to 5.4–30 seconds with V1." That is up to roughly 15x faster for the largest images.

The new runtime keeps AgentCore's existing serverless model — "no pre-provisioning, scale to zero, hardware-enforced session isolation, and pay only for what you use" — and is available now in five regions: us-east-1, us-east-2, us-west-2, eu-west-1, and ap-northeast-1. Existing customers opt in by setting platformVersion to V2 when creating or updating a runtime.

Context

Amazon Bedrock AgentCore is AWS's managed infrastructure for running AI agents in production — the microVM layer that actually executes an agent's code, tool calls, and long-running sessions, separate from which foundation model the agent calls. AWS has been racing to make that layer cheap and fast enough for real workloads: agents routinely sit idle between tool calls or user turns, and the previous runtime's cold-start times of up to 30 seconds for larger images made bursty, request-driven agent traffic expensive and slow to scale.

This release lands the same week AWS made OpenAI's GPT-6 Astra generally available on Bedrock and added xAI's Grok 4.6, underscoring that AWS's competitive angle in agentic AI is increasingly about the execution layer underneath a wide catalog of third-party models, not a model of its own.

Why it matters

Cold-start latency and idle-memory billing are the kind of unglamorous infrastructure details that determine whether "AI agents in production" is actually economical at scale. A 3x-to-15x cut in cold-start time changes what workloads are practical to run serverless — request-triggered agents, bursty customer-support traffic, or short-lived background tasks that previously ate a multi-second latency tax on every invocation.

It also sharpens AWS's pitch against agent-hosting competitors: rather than compete primarily on which foundation models it offers, AWS is optimizing the commodity layer — compute, memory, isolation — that every agent regardless of model provider has to run on. If the performance numbers hold up under real traffic, this is a meaningful cost and latency edge for any team already committed to Bedrock's agent stack.