AWS this week announced that AgentCore now supports persistent compute sessions lasting up to 14 days, allowing multiple AI agents to share runtime dependencies on the same EC2 instance, with break-even achievable at just 24% CPU utilization. Meanwhile, AWS open-sourced Dogwood, a tool-calling policy language; the MCP header mechanism is pushing agent permission declarations; and Cloudflare launched usage-based Agent Tracing. As agents move from per-call invocation to resident operation...

The way AI agents consume compute is undergoing a fundamental shift. This week, AWS announced that AgentCore supports persistent compute sessions of up to 14 days, enabling multiple AI agents to “co-reside” on a single EC2 instance and share the sessions and dependencies of the same runtime. In effect, agents are moving from on-demand “hourly workers” to long-term resident “tenants.”

he business logic behind this model is straightforward. Industry-average resource utilization has long remained in the 30% to 40% range, while AgentCore offers a clear break-even point: once CPU utilization reaches