AWS made runtime instances generally available in Amazon Bedrock AgentCore on 6 August — a managed compute option for production agents in which sessions persist for up to 14 days. It shipped as GA, not preview.
What is in the box
Python 3.11-3.14 with native code, plus container images; Linux on ARM64 and x86_64; GPU-accelerated instance types available. The launch example is a c7g.2xlarge — 8 vCPUs, 16 GiB. Regions at launch are US East (Ohio, N. Virginia), US West (Oregon), Asia Pacific (Mumbai, Singapore, Sydney, Tokyo) and Europe (Frankfurt, Ireland).
Fourteen days of what
Of session state, not of guaranteed uninterrupted execution. What the duration buys is an agent that keeps its working files and context across a long task instead of rehydrating from storage on every call — plus multiple agents collaborating on one host through a shared persistent file system.
The pricing change is the real change
There is no flat per-agent price. Runtime instances bill at standard EC2 rates plus a management fee for orchestration. An agent that holds a session for days bills like a server that runs for days. That moves agent cost from per-token, where it scales with work done, to per-hour, where it scales with time held.
Alongside, not instead
Runtime instances sit next to AgentCore's existing microVM runtime rather than replacing it. Short, stateless invocations still have a cheaper home; this is for the long-running kind the industry has spent a year promising. The practical decision it forces on teams is new — not which model to call, but how long an agent is allowed to sit holding a machine while it waits for something.