Amazon Bedrock AgentCore is AWS's enterprise-grade infrastructure layer for deploying and managing AI agents at scale. It includes a serverless runtime, short- and long-term memory, an MCP-compatible gateway, identity management, a sandboxed code interpreter using Firecracker microVMs, a remote browser, observability via OpenTelemetry, an evaluation service, and a policy engine that enforces business rules outside the agent. The platform is model- and framework-agnostic, supporting Anthropic, OpenAI, Google Gemini, LangGraph, CrewAI, and others. A hands-on walkthrough of the customer support agent demo reveals multiple bugs and documentation issues that required AWS engineer fixes, though they were resolved quickly. A notable design concern is that the external policy module blocks agent actions without passing the denial reason back to the model, causing the LLM to hallucinate explanations. Despite rough edges in the samples, the core services appear solid and the total demo cost was $0.35.
Sort: