The new AgentCore runtime: Elastic, optimized, and consistently fast starts

AWS announces the new AgentCore runtime, a managed compute layer for deploying and running AI agents with improved performance and cost efficiency.

• Pages memory on demand and reclaims unused memory instead of holding peak allocation throughout session
• Delivers consistent cold start times (~2 seconds) regardless of container image size or concurrency
• Snapshots agent initialization once, then restores for each new instance to eliminate repeated startup work
• Consumption-based billing tracks actual memory usage rather than provisioned capacity
• Upcoming features include committed baseline discounts, larger compute/storage, x86 support, and scoped identity for unattended agents

The new runtime optimizes infrastructure for production agents running longer, more complex workloads while maintaining serverless scale-to-zero economics.