What changed
AWS made Amazon Bedrock AgentCore runtime instances generally available, letting teams place agents on managed EC2 capacity providers with GPU, memory-optimized or compute-optimized instance types while AWS handles provisioning, patching, scaling and lifecycle management. Sessions can run for up to 14 days, versus up to eight hours for the serverless microVM path.
Why it matters
Long-running or specialized agents now have a managed path to dedicated EC2 shapes without changing the invoke model. The choice also creates a distinct capacity, tenancy and cost boundary that must be compared with the serverless runtime rather than treated as a transparent upgrade.
What to do
Deploy one resumable synthetic job through both runtime types. Record cold start, session duration, interruption and recovery, instance placement, patching, logs, network and identity boundaries, GPU fit, billed management and EC2 cost, then terminate the capacity provider and verify cleanup before choosing a default.
Caveat
General availability and the 14-day limit are AWS evidence, not independent reliability or isolation results. Runtime instances add AgentCore management charges to EC2 costs, support only named regions, and still require account quotas, capacity, networking, image, credential and failure-recovery testing.