News

AWS ships next-generation AgentCore runtime with sub-2-second starts and idle scale-to-zero

Published 3 min readAI-assisted · human approved
TL;DR

AWS announced on 12 November an updated Amazon Bedrock AgentCore runtime that it says starts sessions in under 2 seconds, offers up to 8 vCPUs and 32 GB, and bills less for idle agents.

Why it matters

If you run agents that execute code or wait on tools, startup time, memory billing and where data lives decide both cost and security design. These are AWS's own figures, so check them against your own workloads.

Image from AWS Machine Learning Blog
Image: AWS Machine Learning Blog. It belongs to its owner; see our image policy.

Original source

AWS announced a next-generation version of the AgentCore runtime in Amazon Bedrock AgentCore on 12 November 2026. The managed compute layer keeps its one-MicroVM-per-session isolation and adds larger sessions, faster startup, usage-based memory billing and filesystem mounting.

What happened

In a post on the AWS Machine Learning Blog, Madhu Parthasarathy describes updates to AgentCore runtime, the managed environment for deploying and running agents without operating the infrastructure yourself. AWS says every session still runs in its own hardware-isolated Firecracker micro virtual machine (MicroVM), with its own identity, policy and audit trail.

AWS's headline figures are up to 15x quicker session starts, 4x the compute and up to 68% better efficiency. These are the company's claims, and the post pairs them with customer quotes rather than an independent benchmark.

Key details

  • Bigger sessions: sessions scale up to 8 vCPUs and 32 GB, with container images up to 10 GB. AWS describes this as 4x the previous version.
  • Startup: sessions start in under 2 seconds, independent of image size and concurrency. AWS says you no longer need warm pools or pre-provisioned clusters to manage cold starts.
  • Idle handling: the runtime reclaims memory within minutes of an agent releasing it, and an idle agent scales to zero. For multi-turn agents with natural pauses, AWS says this means up to 68 percent less memory per session.
  • Committed baseline pricing: for always-on agents, you set a memory baseline per session that is billed as a minimum, with room to burst up to four times that baseline on demand. AWS says this gives steady workloads a predictable rate at AWS Lambda MicroVM rates.
  • Data inside the boundary: agents can dynamically mount a filesystem and save their work to it, for example an Amazon S3 bucket with FUSE support. AWS says the data stays inside the trust boundary for the whole job, under one set of controls defined once.

What customers say

The post includes quotes from adopters. They are testimonials chosen by AWS, not audited results.

Customer Reported result
Mission Cloud In an initial comparison of both runtime versions: 35% faster cold starts for inference requests, 77% faster starts for health checks, 17% lower measured runtime cost per call
Zuora Memory footprint per invocation cut by around 60% in many use cases, with more traffic handled at steady latency

Coinbase, Okta, Rescale, Deloitte Consulting and Caylent also provided quotes. They focus on isolated coding-agent sessions, safe scaling of engineering agents, and cost and latency reductions.

Why it matters

Two parts of the update speak directly to builders.

First, agents spend much of their time waiting on a model response, a tool call or a person. Billing that follows memory actually held, plus scale-to-zero when idle, changes the cost model for chat-style or approval-driven agents. Always-on agents get a different option: a minimum baseline with burst headroom.

Second, the filesystem mounting addresses a common design problem. An agent may be isolated, yet the files and context it needs sit outside that boundary, so teams write their own rules for moving data in and out. Mounting storage into the session is AWS's answer. Possible uses include coding agents that compile large codebases or run test suites against a mounted workspace, though the post's examples of workloads are broader (browser-driven workflows, local model inference).

Before migrating, measure your own agents. The Mission Cloud and Zuora numbers come from specific workloads, and the 68% memory figure is described for multi-turn agents with pauses, not for every pattern. For related AWS agent work, see how AWS evaluates agents with UAEF and a serverless MCP server for data lake access on ECS Fargate.

Checked claims

ClaimStatusEvidence
Each AgentCore runtime session runs in its own hardware-isolated Firecracker MicroVM, according to AWS.confirmedevery session runs in its own hardware-isolated Firecracker micro virtual machine (MicroVM)
Sessions scale to up to 8 vCPUs and 32 GB, with container images up to 10 GB.confirmedup to 8 vCPUs and 32 GB, with support for container images up to 10 GB
AWS says sessions start in under 2 seconds regardless of image size and concurrency.confirmedSessions start in under 2 seconds, independent of image size and concurrency.
Idle agents scale to zero under the updated runtime.confirmedan idle agent scales to zero
Agents can mount a filesystem such as an S3 bucket with FUSE support to save work inside the trust boundary.confirmedAgents can now dynamically mount and save their work to a filesystem
Mission Cloud reported faster cold starts and lower per-call runtime cost in an initial comparison of the two runtime versions.confirmedthe same workload delivered 35% faster cold starts for inference requests and 77% faster starts for health checks, while reducing measured runtime cost per call by 17%
0
#aws#agentcore#agents#amazon-bedrock

Found a mistake? Use Report above; an editor reads every one. Editorial policy

Question for readersWhich of your agents spend most of their time idle, waiting on models, tools or people, and how would usage-based memory billing change what you run?

Discussion · 0

No comments yet. Start the discussion.

Sign in to join the discussion.