The new AgentCore runtime: Elastic, optimized, and consistently fast starts
Amazon Web Services Evandro Franco ● Covered by 2 sources
AWS says its new AgentCore runtime keeps agent startups fast and memory use lower. That matters because production agents now run for hours, not just chat for a minute.
Based on reporting by Amazon Web Services, Evandro Franco — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is updating the runtime under Amazon Bedrock AgentCore, its managed layer for running agents without having to build or babysit infrastructure. The pitch is simple: agents have grown up. They’re no longer just chat toys; they write code, coordinate systems, and keep working long after a human has walked away.
The new runtime is built around two pain points AWS says show up in real production use. First, memory. The old model held onto a session’s peak usage until the session ended, which is fine for short exchanges and annoying for long, bursty jobs. The new one starts from a smaller memory profile, adds more only when needed, and reclaims memory when it goes cold.
Second, cold starts. AWS says startup latency used to swing with image size and concurrency, which is exactly the kind of thing users notice when an agent pauses for input and then has to resume. The new runtime snapshots the environment after it becomes healthy, then restores that snapshot for each new instance. In AWS’s test setup, that produced a P75 cold start of about 2 seconds across image sizes from 200 MB to 2 GB, while the original runtime rose from roughly 5.4 seconds to nearly 30 seconds.
AWS framed the test carefully: an empty echo agent, 5,000 cold invocations per agent, five image sizes, no model calls, no tools. The point was to isolate the runtime itself. And in that setup, the agent code took about 34 milliseconds at P75, so almost all of the waiting came from the platform start path.
The company is also betting that the economics improve with the technical changes. Billing now follows the memory actually used during the session instead of charging against the full resident footprint. AWS says the rate is higher, but the number of GB-hours falls enough that most agents should pay less overall. For teams trying to keep a lot of agents idle most of the time, that’s the whole game.
My take — AI-written commentary, not fact-checked reporting
This is the sort of infrastructure update that matters more than another flashy agent demo. Everyone wants autonomous agents; fewer people want to pay for peak memory and warm capacity just to keep them from stumbling at startup. AWS is basically admitting the old cloud trick of “keep some spare stuff around, trust me” doesn’t fit agent workloads very well anymore.
Read more about this at: Amazon Web Services