AI agents are evolving beyond single commands to running multi-hour, complex workflows autonomously. This increased longevity and autonomy highlight the challenge of securely, reliably, and economically operating agents at scale within enterprises, especially as 40% are scaling AI agents, up from 27% a year prior.
Amazon Bedrock's AgentCore runtime provides a fully managed compute environment using hardware-isolated Firecracker micro virtual machines, ensuring each agent session has strong isolation, identity, policy enforcement, and an audit trail. This isolation is crucial because agents execute code, interact with browsers, access sensitive data, and act on users’ behalf, raising security demands beyond traditional software.
The updated AgentCore runtime can handle sessions up to four times larger—supporting up to 8 virtual CPUs and 32GB memory—and start sessions in under two seconds regardless of container image size. This eliminates the need for pre-warmed environments, reducing operational complexity and latency in scaling multi-agent deployments.
Resource efficiency and cost control are improved with dynamic memory reclamation and scaling to zero memory use for idle agents, lowering memory consumption per session by up to 68%. The billing model aligns charges with active usage rather than idle time, offering financially predictable and optimized operation for enterprises moving agents into production.
Unique to this generation, AgentCore extends isolation beyond compute to data by letting agents dynamically mount and save files within trusted boundaries, such as Amazon S3 buckets. This integration simplifies governance by unifying access controls and audit trails around both code execution and the data accessed, reducing complexity and risk.
Customer testimonials from leading firms like Coinbase, Okta, and Thomson Reuters confirm that these improvements enable faster agent start times, enhanced reliability, reduced costs, and secure scaling without managing infrastructure, facilitating production deployment of AI-powered workflows.
Despite advancements, scaling AI agents securely remains complex. Only a minority of organizations have fully operational multi-agent systems, underlining the ongoing challenge of trust and governance in autonomous AI. Amazon’s updates represent important infrastructure progress but do not eliminate the need for careful policy and security management.
The next generation of AgentCore runtime is broadly available across AWS regions, aiming to provide organizations with a scalable, secure, efficient foundation that prioritizes isolation and operational excellence as AI agents become integral to business processes.
Source & further reading
Run agents securely and efficiently with the next generation AgentCore runtime ↗Source published 2026-11-12. Claims and limitations should be checked against the original announcement or research.