AWS AgentCore V2 Cuts Cold Starts to 2 Seconds Across All Image Sizes

AWS shipped the new AgentCore runtime on 18 September 2026, targeting two failure modes that push production agent teams toward expensive self-managed workarounds: unpredictable cold-start latency that scales with image size, and memory billing anchored to session peaks rather than actual consumption. Both problems grow more acute as agents shift from short interactive turns toward ambient, event-driven workloads that run for hours without supervision. As we've argued before, production AI fails on architecture, not model intelligence — and this release is a direct response to that reality.

Dynamic memory reclamation replaces peak-hold billing

The original AgentCore runtime allocated memory at session start and held it at the high-water mark until the session terminated. For bursty or long-running agents that spike during tool calls and then sit idle, the billing gap between peak allocation and actual working-set usage could be substantial. The new runtime inverts this: it begins each session from a small resident footprint, pages in additional memory as the workload touches it, and actively reclaims memory that has been released or gone cold rather than waiting for session end. AWS states this approach is tuned from analysis of allocation patterns across billions of sessions.

The pricing trade-off is deliberate. The per-GB-hour rate for V2 is higher than V1, but the billable footprint shrinks because reclaimed memory drops out of the meter in real time. AWS's position is that for most agents the footprint reduction outpaces the rate increase, producing a lower total bill. Teams with flat, memory-intensive workloads that genuinely sustain high utilization throughout a session will need to validate this against their own allocation profiles before migrating.

Snapshot-based cold starts decouple latency from image size

When an operator creates or updates a V2 runtime, AgentCore launches the container, waits for a healthy signal, and captures a snapshot of the running environment — after one-time initialization work such as loading model artifacts and fetching static configuration has already completed. Every subsequent instance restores from that snapshot rather than repeating the boot-and-initialize sequence from scratch.

The critical engineering detail is how AWS keeps snapshot size flat as container images grow. A naïve snapshot of a running process includes caches and transient allocations that inflate restore size and cause latency to scale with image size. V2 strips that excess, retaining only the working state required to resume execution. The result: snapshot size stays roughly constant across a wide range of image sizes, which is what holds restore latency steady.

AWS measured cold-start performance by sending 5,000 cold invocations per agent to an empty echo agent (no model calls, no tool calls) from a Python boto3 client on an EC2 instance in us-west-2, targeting us-east-1 over the public internet without VPC peering. The figures therefore include cross-region round-trip time on top of platform start time. The agent's own code executed in approximately 34 milliseconds at P75 in the echo test, so the cold-start numbers are almost entirely platform overhead.

A session that lands on an already-initialized environment starts in under 100 milliseconds on either version; the snapshot mechanism specifically targets cold paths. AWS also notes that interactive agents can hide the remaining 2-second cold start by initiating the session the moment a user opens the interface rather than waiting for first input submission.

What's coming in the V2 roadmap

Several capabilities are listed as forthcoming. Baseline pricing will let operators reserve a memory floor for steady, always-active sessions and burst above it on demand, complementing the existing consumption model which remains available for spiky or scale-to-zero workloads. Larger compute and storage configurations, x86 microVM support (current V2 runs non-x86), suspend-and-resume with memory snapshotting, and scoped per-session identity for unattended agents are all described as in progress. The identity feature is particularly relevant to infrastructure governance for safe agent deployment: session context keys that constrain an unattended agent to exactly its defined permissions address a real gap that chat-turn agents never exposed.

To activate V2, operators set platformVersion to V2 on create or update calls; the change is API-driven with no infrastructure provisioning required.

AI Mastery analysis

The snapshot approach is well-understood in serverless compute — AWS Lambda's SnapStart for JVM workloads uses the same restore-from-checkpoint pattern — but applying it to agent runtimes introduces a constraint worth flagging: any state captured in the snapshot is shared across all instances restored from it. Secrets, per-user tokens, or mutable configuration fetched during initialization will be present in every restored instance. Teams that currently initialize agent sessions with runtime-specific credentials must ensure those are injected post-restore, not pre-snapshot, or they risk credential leakage across session boundaries. This is the same class of infrastructure isolation problem that has surfaced repeatedly as agent workloads have scaled.

The memory reclamation model also assumes that agent frameworks cooperate by actually releasing memory between requests. Frameworks that cache aggressively — retaining large context windows, embedding caches, or tool-call result buffers — will see less benefit because the runtime can only reclaim what the process has freed. Developers should profile their framework's allocation behavior under V1 before assuming V2 delivers the advertised footprint reduction. The ~2-second P75 cold start figure is cross-region with public internet routing included; same-region, VPC-internal deployments will likely read lower, and that number should be established experimentally before being used in SLA commitments.

This release continues the pattern of infrastructure-layer rewrites driving more measurable gains than model-weight improvements for teams already in production. The snapshot-based cold start is the more durable advance: removing image size as a latency variable simplifies the entire container optimization conversation for agent teams and removes a significant operational incentive to keep images artificially small at the cost of missing dependencies.

Primary source

The new AgentCore runtime: Elastic, optimized, and consistently fast starts — AWS Machine Learning Blog