AWS has announced a new generation of its Amazon Bedrock AgentCore runtime, with sessions that start in under two seconds, scale to four times more compute and can mount storage inside the session’s trust boundary. For teams running AI agents, the changes target three practical headaches at once: slow starts, idle resource costs and keeping data governed while an agent works.
AWS AI Watch analysis
What happened
AWS says each session continues to run in its own hardware-isolated Firecracker micro virtual machine, with its own identity, policy and audit trail. The new runtime supports up to 8 vCPUs and 32 GB of memory, plus container images up to 10 GB. AWS says sessions start in under two seconds regardless of image size or concurrency.
The runtime can reclaim memory within minutes when an agent releases it, and idle agents can scale to zero. For multi-turn agents with pauses between interactions, AWS says this can reduce memory consumed per session by up to 68%. Agents can also mount and save work to a filesystem, including an Amazon S3 bucket with FUSE support, while data remains within the session’s trust boundary. AWS’s announcement says the new runtime is available today in regions across North America, South America, Europe and Asia Pacific.
Why it matters
A slow cold start can make an agent feel sluggish; keeping spare environments warm to avoid that delay can cost money while they sit idle. AWS’s changes aim to reduce both the wait and the waste, while giving teams more room to run demanding workloads such as code compilation, browser workflows and local model inference.
The storage change addresses a different operational wrinkle: an isolated agent still needs access to files and working context. AWS says teams can govern mounted data under the session’s existing controls rather than building separate rules for how it moves in and out. That is a concrete infrastructure change, not a guarantee that every agent deployment is secure or cheaper.
Our read
This is a substantial upgrade for people building agents that need to run real workloads, not just complete a tidy demo. Faster starts, more capacity, scale-to-zero behaviour and governed storage make a useful package. Operators should test it against their own workloads and costs; AWS’s headline figures are its claims, not a promise that every deployment will see the same gains.
What to watch
- Whether teams see the promised start times under their actual workloads and concurrency.
- How memory reclamation and scale-to-zero affect cost and responsiveness in practice.
- Whether mounted storage simplifies governance without adding friction for agent developers.
Discussion spark: For production AI agents, which matters most: faster starts, lower idle costs, or keeping working data inside the session’s security boundary?
Sources and evidence
- Run agents securely and efficiently with the next generation AgentCore runtime (12 November 2026, 17:59 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)