Amazon Web Services has made the second generation of its Bedrock AgentCore runtime the default for new agent deployments, a change announced on November 12 that should make AI agents start far faster — and cost less when they sit idle. The new runtime restores each agent session from a snapshot of an already-initialised agent, instead of booting a container and setting the agent up from scratch every time.

Cold starts fall from 29.7 seconds to under two

AWS's own numbers, using a 2 GB container image, put 75 percent of sessions (P75) starting in under two seconds, compared with 29.7 seconds on the previous generation. AWS describes the improvement as up to 15 times faster. One customer migrating to the new runtime reported cold starts falling from nearly a minute to under five seconds, even during peak traffic. V2 first shipped as an opt-in platform version on September 18; from November 12 it is the default for newly created runtimes in regions across North America, South America, Europe and Asia Pacific, including Northern Virginia, Ohio, Oregon, Frankfurt, Ireland, London, Tokyo, Seoul and Sydney.

Bigger sessions, cheaper idle time

Each session can now use up to 8 vCPUs and 32 GB of memory — four times the compute of the previous version — and container images can grow to 10 GB. Idle agents scale down to zero, and memory an agent is not using is reclaimed after 120 seconds instead of being held until the session closes. AWS says multi-turn sessions can use up to 68 percent less billed memory as a result. There is a trade-off: V2's list rates are higher than V1's — $0.1276 per vCPU-hour and $0.0169 per GB-hour, against $0.0895 and $0.00945 — so the savings depend on how much idle time the new memory handling removes. A committed-baseline billing option is also available.

Existing deployments stay put — for now

The change applies only to newly created runtimes. Existing deployments stay on their current version until customers move them, and developers who want to stay on the old generation can still select platformVersion V1 when creating a runtime. Each session continues to run in its own Firecracker microVM, with hardware-level isolation, its own identity, access policy and audit trail. AgentCore Runtime is the managed compute layer of Bedrock AgentCore, AWS's platform for running agents built with different frameworks and models — the part that hosts the agent so developers do not manage servers themselves. The move matters because cold starts have been one of the quiet taxes on agentic software: every time a user opens a chat or triggers a workflow, the agent has to wake up before it can do anything useful. Cutting that wait from half a minute to under two seconds makes agents feel instant — and, with idle costs falling too, makes always-on agents cheaper to leave running.