Amazon Launches AgentCore Framework
Amazon introduced AgentCore, a new framework on its Bedrock platform designed to standardize the development and deployment of AI agents. The framework provides a unified architecture for building conversational AI, integrating knowledge bases, and defining agent capabilities. Amazon claims AgentCore can reduce infrastructure costs by up to 60% and shorten development time.
- The framework is a modular suite of services, not a monolith, including a serverless Runtime, a Gateway to connect to APIs like Slack or Salesforce, managed Memory, and an Identity service that integrates with providers like Okta and Microsoft Entra ID. - AgentCore is designed to be framework-agnostic, acting as a production layer for agents built with popular open-source libraries like LangGraph, CrewAI, and LlamaIndex, rather than being a direct competitor to them. - The cost reduction is achieved through a consumption-based pricing model for the AgentCore Runtime; customers are not charged for idle CPU time during I/O wait periods, which can account for 30-70% of an agent's session time while waiting for LLM or API responses. - It provides built-in, sandboxed tools including a Code Interpreter for executing data analysis or mathematical tasks and a secure Browser for automating web-based workflows. - For workflow orchestration, AgentCore has native integration with AWS Step Functions, allowing developers to build, visually inspect, and audit complex, multi-step agentic processes without writing custom Lambda invocation code. - The AgentCore Memory component provides both short-term memory for conversational context within a single session and long-term memory that persists across multiple interactions to enable learning. - For monitoring, the AgentCore Observability service provides real-time dashboards in Amazon CloudWatch and supports OpenTelemetry (OTel), allowing for integration with existing monitoring systems. - New capabilities in preview include "Policy," which allows teams to define agent boundaries and block unauthorized actions in real-time, and "Evaluations" for continuously inspecting agent quality on dimensions like correctness and goal success rate.
