AWS makes Bedrock AgentCore runtime instances GA for production AI agents

AWS made runtime instances generally available in Amazon Bedrock AgentCore, offering persistent, managed EC2-based compute for production AI agents. The new option supports multi-agent collaboration, GPU workloads, and long-running sessions of up to 14 days, complementing the existing lighter-weight microVM-based runtime. It targets teams that need durable, stateful agent infrastructure rather than ephemeral function-style execution.

AWS anchored the launch with a customer proof point: Mobileye built a production-grade AI Support Agent on AgentCore that saves engineers hundreds of hours on routine ticket lookups. VP of Agentic AI Swami Sivasubramanian has been publicly championing the agent stack, and AWS separately open-sourced Kiro Crew (originally an internal project called Meshclaw) used by tens of thousands of Amazon builders.

The runtime GA is one piece of a broad AgentCore push this week that also included built-in web search for foundation-model grounding, an n8n integration harness, and a suite of governance controls (covered separately). Together they signal AWS's strategy: rather than compete head-on at the frontier model layer, own the production infrastructure and governance plane where enterprises actually deploy agents.

Competitively, this pits AWS against Azure's Microsoft Agent Framework and Google's managed agents with remote MCP — all three hyperscalers shipped agent-runtime and orchestration capabilities the same week, marking agent infrastructure as a clear theme. The differentiator AWS is pressing is persistence and session length: 14-day sessions and GPU support target the most demanding long-horizon agentic workloads. Watch whether enterprises adopt the heavier runtime or stick with cheaper microVMs for most use cases, and how pricing compares as agents run for days rather than seconds.