AI agents have moved past one-off prompts. They now run multi-step workloads for hours at a time. The length of task that frontier agents can complete on their own doubles about every four months (METR), and 40 percent of large enterprises are scaling AI agents, up from 27 percent a year ago (McKinsey). As agents run longer and with more autonomy, the focus shifts from what agents can do to how securely, reliably, and economically they can operate at scale. Agents do more than traditional software: they execute code, navigate browsers, access sensitive data, and act on behalf of users. These capabilities introduce new security and infrastructure demands, making strong isolation essential for production deployments.
Every agent session needs a secure, isolated environment. But delivering that isolation at scale introduces a fundamental challenge: balancing performance, cost, and efficiency. Pre-provisioned environments consume resources even when idle, driving up costs. On-demand environments reduce wasted capacity but introduce cold-start latency, delaying agent execution. Neither approach scales efficiently without trade-offs. As organizations move from deploying a handful of agents to running thousands of concurrent sessions, building and managing custom sandboxes becomes increasingly complex, expensive, and unpredictable.
The challenge is evident. Only 15 percent of organizations have scaled orchestrated, multi-agent systems, and 70 percent say they can’t yet trust and govern their agents (Deloitte). Scaling agents shouldn’t mean choosing between security, speed, and cost. The next generation of agent infrastructure must deliver strong isolation with fast startup times, elastic scaling, and usage-based economics. Because as AI agents move into production, secure execution shouldn’t be the bottleneck. It should be the foundation that makes scale possible. The fastest path to production should also be the safest one.
The next generation AgentCore runtime
We built Amazon Bedrock AgentCore to help developers build, connect, and optimize agents securely at scale. AgentCore runtime, a capability of Amazon Bedrock AgentCore, is the managed compute layer. It gives developers a fully managed environment to deploy and run agents without building or maintaining infrastructure. Since launching, our customers have been pushing their agents into increasingly complex workloads. As those workloads grow, so do the demands on the execution environment: more compute, more flexible billing, and trust boundaries that extend to data, not only compute. Agents reach production at scale if the infrastructure underneath was built for both scale and speed.
Today we’re announcing updates to AgentCore runtime that meet those demands. The core guarantee stands: every session runs in its own hardware-isolated Firecracker micro virtual machine (MicroVM), with its own identity, policy, and audit trail. These updates expand what’s possible within that guarantee. Sessions now start up to 15x faster, scale to 4x more compute, and run up to 68% more efficiently, with the isolation boundary fully intact.
Here’s what’s new.
Run bigger workloads without waiting longer to start
Teams building increasingly advanced agents asked for more flexibility: the ability to scale memory and compute within the isolated session to match the demands of their workloads. From large codebase compilations to browser-driven workflows to local model inference, the execution environment needs to keep pace.
Sessions now scale 4x compared to the previous version, up to 8 vCPUs and 32 GB, with support for container images up to 10 GB. Sessions start in under 2 seconds, independent of image size and concurrency. You no longer need warm pools and pre-provisioned clusters to manage cold starts.
“Migrating to the next generation of AgentCore runtime reduced our agent session cold starts from nearly a minute to under five seconds, even during peak traffic. It also eliminated startup delays longer than 30 seconds, all while our weekly agent volume more than doubled.”
— Siwei Shen, VP of Engineering, Coinbase
“Mission is using Amazon Bedrock AgentCore runtime to power a natural-language query agent that helps users retrieve property and sales insights without navigating complex data systems. In our initial comparison of both runtime versions, the 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%. The new AgentCore runtime gives us a compelling combination of faster agent experiences and more efficient resource utilization as we bring enterprise agents into production.”
— Dr. Ryan Ries, Chief AI & Data Scientist, Mission Cloud
“Our approach is to run Zuora’s cloud with automation and AI agents wherever we can, and that only works if the runtime is efficient and reliable. The next generation AgentCore runtime cut our memory footprint per invocation by around 60% in many use cases while handling more traffic at steady latency. That gives us a stronger cost profile as we expand agentic workloads across our platform.”
— Brad Ruppert, VP of Cloud and Security, Zuora
“The Rescale platform helps R&D and engineering teams accelerate product development with agentic digital engineering. Our engineering-specific agents validate simulation inputs, troubleshoot jobs, select hardware, and generate reports, while engineers stay in control. Amazon Bedrock AgentCore runtime allows us to deploy and scale those agents in production without building or operating the underlying agent infrastructure, so we can deliver new agents to customers faster.”
— Mark Whitney, VP Engineering, Rescale
Stop paying for idle agents
As teams move agents closer to production, they asked for billing that reflects how agents actually work. Agents spend much of their time waiting on a model response, a tool call, or a human decision. Agents should be billed for the work they do, not the time they spend waiting.
The runtime now reclaims memory within minutes of the agent releasing it, and an idle agent scales to zero. For multi-turn agents with natural pauses between interactions, this means up to 68 percent less memory consumed per session. For always-on agents, committed baseline pricing sets a memory baseline per session, billed as a minimum, with room to burst up to four times that baseline on demand. This gives steady workloads a predictable rate at AWS Lambda MicroVM rates. The runtime uses compute more efficiently, so you pay less for the same work. That makes the finance conversation that often gates production rollouts easier to close.
“The next generation Amazon Bedrock AgentCore runtime helps us deliver more responsive AI experiences by starting agents faster and using memory more efficiently. Aligning consumption with actual usage gives our clients great confidence to scale agentic solutions while keeping costs connected to real demand.”
— Ashwin Patil, Principal, AI and Data Leader, Deloitte Consulting LLP
“AgentCore runtime gives us a managed foundation for building agentic products and platforms around our customers’ data, tools, and business processes. The next generation Agentcore runtime makes that foundation more responsive and resource-efficient by restoring initialized environments and reclaiming unused memory. We’re beginning to build on the latest version because the cost and latency reductions make a significant difference in how our customers’ businesses operate.”
— Randall Hunt, CTO, Caylent
Isolation now extends to your data
Many enterprises face a persistent gap in their agent security model. The agent runs inside an isolated session, but the data it requires resides outside that boundary. Long-running agents pick up context as they go: files, credentials, tool outputs, and more. Teams end up building their own rules for how data moves in and out, adding complexity and risk with every new agent.
Isolation that stops at compute is only half a boundary. The next generation AgentCore runtime closes that gap. Each session still runs in its own hardware-isolated environment with its own identity, access policy, and audit trail. Agents can now dynamically mount and save their work to a filesystem, such as an Amazon Simple Storage Service (Amazon S3) bucket with FUSE support. Your data stays inside the trust boundary for the entire job, governed by a single set of controls you define once.
“At Okta, the combination of AgentCore runtime and Lambda MicroVMs provide the isolated execution environment and contextual memory our agents need to inspect code, run test suites, and open pull requests within ephemeral sandboxes. Paired with our identity-governed policy controls, AgentCore Runtime and Lambda MicroVMs demonstrate how we can safely scale autonomous engineering agents without risking workspace leakage or privilege escalation.”
— Monica Bajaj, SVP Engineering, Okta
“From helping car shoppers value a trade-in to helping dealers book service appointments and structure deals, Amazon Bedrock AgentCore powers dozens of agentic solutions across Cox Automotive’s brands. Having a single, governed agent platform lets our teams across our business move from pilot to production with confidence.”
— Chris Dillon, VP, Product and Engineering, Cox Automotive
Our commitment
The agents that reach production are the ones you trust to act on your behalf, with real data, real stakes, and real accountability. Trust isn’t a feature you add later. It’s the runtime you build on. This release is a step toward that future: bigger sessions, efficient resource use, and storage inside the trust boundary, without loosening the isolation that makes the work trustworthy. By re:Invent, per-session scoping will bind each session to its own storage folder through session-labeled credentials, and we will keep building from there.
Get started
The next generation AgentCore runtime is available today in AWS Regions across North America, South America, Europe, and Asia Pacific. Visit Amazon Bedrock AgentCore to build on it or open the Amazon Bedrock console to start exploring.
About the author
