Amazon Reworks AgentCore Runtime for Faster AI Agents

Amazon is updating the execution layer behind its Bedrock AgentCore platform with a runtime designed for agents that run longer, wake up unpredictably, and operate without a person watching every step. In an AWS announcement, the company says the new runtime reclaims memory as sessions go cold and keeps cold-start latency consistent across container sizes and concurrency levels.

What changed

The previous model held allocated memory until a session ended. That is tolerable for short conversations, but expensive for agents that run for hours, spike briefly, then wait. The new runtime starts from a smaller memory profile, pages in additional resources as needed, and reclaims memory that is released or unlikely to be accessed again. Amazon says billing therefore follows changing usage instead of the session’s highest memory watermark.

Cold starts are the second target. AWS says it prepares the environment once, snapshots it, and restores that snapshot for new instances. In its test using an echo agent, the new runtime delivered a P75 cold start of about two seconds for images from 200 MB to 2 GB. The previous runtime ranged from roughly 5.4 seconds to nearly 30 seconds as image size increased. Those figures are AWS measurements, not an independent benchmark, but they point at a real production nuisance: infrastructure latency can dominate the user experience before an agent makes its first model call.

Why builders should care

This is more than a serverless tuning release. Agent workloads are shifting from request-response chat toward event-triggered, ambient, and unattended execution. That changes the economics of orchestration: idle time, resumability, memory persistence, and identity boundaries matter as much as model latency.

For teams building workflows in n8n or similar systems, the practical lesson is to measure the entire agent loop, not only token speed. Track wake-up latency, memory retention, tool-call duration, retries, and the cost of sessions that spend most of their life waiting. AWS also says scoped identities for unattended agents are coming, which could make long-running automation safer when each session receives only the permissions required for its task.

The shift fits the broader infrastructure trend covered in n8n Lab’s analysis of Huawei’s agentic cloud push: production agents are becoming an infrastructure problem before they become a prompt problem.