Amazon Bedrock's AgentCore harness enables serverless image editing with natural language

Amazon Bedrock has released AgentCore harness, a configuration-driven framework for building serverless AI agents that orchestrate complex workflows without custom orchestration code. This post demonstrates an image editing agent where users upload a photo, describe an edit like “change the car color to blue,” and receive the result in seconds. The agent, powered by Claude Sonnet 4.6 and Claude Haiku 4.5, routes tasks to three Stability AI-based tools for inpainting, outpainting, and search-and-replace. The entire stack—including authentication, encrypted storage, a React frontend, and the agent itself—deploys with a single AWS CDK command. AgentCore runs each invocation in a stateful, isolated microVM with built-in memory (30-day persistence), tool routing via Model Context Protocol (MCP), and a post-processing step to add watermarks using a shell command (no token cost).

Key capabilities include per-invocation model switching, where the agent preserves conversation context when switching between Claude Haiku for chat and Claude Sonnet for editing, and persona overrides (Real Estate, Retail, Automotive) that inject domain-specific system prompts without redeploying. The agent definition is a simple API parameter set: no Python loops, no streaming handlers, no error retry logic. All tools are exposed as AWS Lambda functions behind an MCP gateway, and the agent selects the correct tool based solely on the prompt. The sample also preserves session IDs in localStorage, so conversations survive browser refreshes. This approach drastically reduces the complexity of building multi-step AI applications, moving from boilerplate orchestration to declarative configuration.