Amazon Bedrock AgentCore Runtime Instances Reach GA — Persistent Compute for Production Agents

AWS announced general availability of runtime instances in Amazon Bedrock AgentCore: persistent, managed EC2 infrastructure purpose-built for production AI agents. Key capabilities include multi-agent collaboration, GPU support, and sessions lasting up to 14 days — a major step beyond ephemeral request-response inference, aimed at long-running autonomous agents that need to maintain state and compute over extended tasks.

The launch is part of a broad AgentCore expansion this week. AWS also brought AgentCore memory, policy, and harness features to AWS GovCloud (US-West) for regulated workloads, letting government and compliance-bound teams build context-aware agents. AgentCore Memory provides persistent context, while policy and harness controls help teams scale from prototype to production.

AWS VP Swami Sivasubramanian highlighted additional new controls: temporal policies powered by Dogwood, a new open-source policy language for AI agents, plus rate limiting in the gateway. His framing captured the emerging problem: 'Most agent failures aren't the result of a single forbidden action, but a sequence of permitted ones' — an agent placing many individually-legitimate purchase orders that collectively blow the budget. Temporal policies close that gap by reasoning about sequences of actions over time.

Real-world adoption is already visible: Cohere Health built a multi-tenant agentic architecture on AgentCore's MicroVM isolation to digitize clinical policies, and Georgia State's TReNDS cut root-cause analysis from 15-30 minutes to under 60 seconds using AgentCore and the open-source Strands Agents SDK. Competitively, AWS is racing Google's Vertex agent platform and Microsoft's Copilot agent stack — the differentiator here is production-grade persistence and governance, betting enterprises care most about controlling agents at scale.