AWS Expands Bedrock with Long-Running Agent Compute and Strands Harness

AWS announced major architectural updates across Amazon Bedrock, spotlighting dedicated AgentCore runtime instances and the open-sourcing of the Strands Agent Harness SDK. The new AgentCore compute instances allow autonomous agents to execute continuous, stateful workflows for up to 14 days on dedicated Amazon EC2 compute—including memory-optimized and GPU-accelerated instances—with regional rollouts in US West (N. California) and Asia Pacific (Hyderabad). Additionally, AWS introduced rate-limiting controls over inference tokens and concurrent user connections, programmatic payment boundaries for tool and MCP resource access, and domain-filtered web search integration. This development matters because production LLM deployments have reached an architectural ceiling with short-lived, request-response execution loops. Engineering organizations trying to build complex agents—such as autonomous coding assistants, multi-step data pipelines, or ongoing site reliability bots—have had to juggle complex external orchestrators, brittle checkpointing databases, and open-ended financial risk. Providing infrastructure-backed session persistence with hard-capped API spending and token rate controls gives DevOps and AI platform teams the enterprise governance required to delegate critical multi-day tasks to autonomous models without risking runaway compute bills or orphaned tasks. In context, this move reinforces the broader AI industry trajectory from raw foundation model evaluation toward holistic runtime environments. Over the past year, frontier models have acquired reasoning and tool-calling capabilities that far exceed traditional serverless invocation timeouts. Cloud providers are racing to provide the surrounding infrastructure plane—the agent harness, secure execution sandboxes, persistent memory fabrics, and deterministic access controls. By coupling native EC2 compute flexibility with Bedrock's foundation model catalog, AWS is attempting to anchor enterprise agentic orchestration directly into its existing cloud governance and IAM perimeter, blunting the pull toward third-party agent hosting frameworks. In practice, practitioners should re-evaluate their agent architecture blueprints. Teams currently using serverless wrappers like AWS Lambda for agent loops should assess whether multi-day AgentCore runtime instances offer better operational simplicity and state persistence for asynchronous workflows. Furthermore, platform architects must immediately implement the new token and connection rate limits alongside payment governance policies, preventing autonomous loops from depleting API budgets during exploratory tool use. Finally, adopting open harnesses like the Strands SDK enables DevOps engineers to maintain workload portability, mitigating provider lock-in while leveraging dedicated cloud compute where necessary.