AWS launched the AgentCore harness in public preview on April 22, 2026, a fully-managed agent runtime that eliminates the need to write orchestration loops, manage state, or wire up memory and observability manually. Instead of packaging agent code into Lambda or Docker containers, developers call a single `InvokeHarnessCommand` API with a model, system prompt, and tools — AWS handles the rest. Key features include per-session microVMs, multi-provider model support (Bedrock, OpenAI, Gemini) with mid-session switching, server-side conversation memory, built-in tools (browser, code interpreter, MCP), and structured observability via CloudWatch and CloudTrail. There is no separate harness charge; you pay only for underlying AgentCore components and model tokens. Current limitations include availability in only four regions, vendor lock-in to AgentCore APIs, incomplete IaC support (Terraform coming soon), and observability tooling still in development.
Table of contents
What is Bedrock, AgentCore, Strands? A quick mapBuilding agents on AWS today still sucksWhat AWS just shippedConfiguring a HarnessThe 20-line agentWhat does this cost?Where it falls shortConclusionSort: